window.LOCAL_FRONTIER_MODEL_DATA = { "schema_version": "2", "generated_at": "2026-07-01T07:10:05.476Z", "assumptions": { "default_overhead_gb": 8, "default_l_alloc_tokens": 100000, "default_l_read_tokens": 32000, "default_r_star_toks": 20, "default_rho": 1, "unit": "decimal GB", "hf_pipeline_filter": "leaderboard-compatible LLM and multimodal candidate filter", "hf_llm_pipeline_tags": [ "text-generation", "image-text-to-text", "video-text-to-text", "audio-text-to-text", "any-to-any", "text-to-audio" ], "min_hf_downloads": 100000, "generated_by": "scripts/generate-model-data.mjs" }, "models": [ { "id": "Qwen/Qwen3-0.6B", "name": "Qwen3-0.6B BF16", "short_name": "Qwen3 0.6B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 27628947, "tags": [ "transformers", "safetensors", "qwen3", "text-generation", "conversational", "arxiv:2505.09388", "base_model:Qwen/Qwen3-0.6B-Base", "base_model:finetune:Qwen/Qwen3-0.6B-Base", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 0.7516, "active_params_b": 0.7516, "layers": 28, "attention_heads": 16, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 1.5032, "active_weight_gb": 1.5032, "kv_alloc_gb_per_1k_tokens": 0.114688, "kv_read_gb_per_1k_tokens": 0.114688, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=27628947; data quality: config, params, base:Qwen/Qwen3-0.6B-Base. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-0.6B" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-0.6B/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-0.6B-Base" } ] }, { "id": "Qwen/Qwen3-4B", "name": "Qwen3-4B BF16", "short_name": "Qwen3 4B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 14472618, "tags": [ "transformers", "safetensors", "text-generation", "arxiv:2309.00071", "arxiv:2505.09388", "base_model:Qwen/Qwen3-4B-Base", "base_model:finetune:Qwen/Qwen3-4B-Base", "license:apache-2.0", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 4.0225, "active_params_b": 4.0225, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 8.045, "active_weight_gb": 8.045, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=14472618; data quality: config, params, base:Qwen/Qwen3-4B-Base. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-4B" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-4B/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-4B-Base" } ] }, { "id": "Qwen/Qwen3-8B", "name": "Qwen3-8B BF16", "short_name": "Qwen3 8B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 13672043, "tags": [ "transformers", "safetensors", "qwen3", "text-generation", "conversational", "arxiv:2309.00071", "arxiv:2505.09388", "base_model:Qwen/Qwen3-8B-Base", "base_model:finetune:Qwen/Qwen3-8B-Base", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 8.1907, "active_params_b": 8.1907, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 16.3814, "active_weight_gb": 16.3814, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=13672043; data quality: config, params, base:Qwen/Qwen3-8B-Base. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-8B" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-8B/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-8B-Base" } ] }, { "id": "google/gemma-4-26B-A4B-it", "name": "gemma-4-26B-A4B-it BF16", "short_name": "gemma 4 26B A4B it", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 13082323, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "conversational", "base_model:google/gemma-4-26B-A4B", "base_model:finetune:google/gemma-4-26B-A4B", "license:apache-2.0", "eval-results", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "MoE transformer", "detail": "gemma4_text", "total_params_b": 26.5441, "active_params_b": 4, "layers": 30, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 1, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 53.0882, "active_weight_gb": 8, "kv_alloc_gb_per_1k_tokens": 0.20992, "kv_read_gb_per_1k_tokens": 0.20992, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=13082323; data quality: config, params, base:google/gemma-4-26B-A4B. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it" }, { "label": "Model config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-26B-A4B" } ] }, { "id": "Qwen/Qwen2.5-7B-Instruct", "name": "Qwen2.5-7B-Instruct BF16", "short_name": "Qwen2.5 7B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 12723410, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "chat", "conversational", "en", "arxiv:2309.00071", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-7B", "base_model:finetune:Qwen/Qwen2.5-7B", "license:apache-2.0", "eval-results", "text-generation-inference", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 7.6156, "active_params_b": 7.6156, "layers": 28, "attention_heads": 28, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 15.2312, "active_weight_gb": 15.2312, "kv_alloc_gb_per_1k_tokens": 0.057344, "kv_read_gb_per_1k_tokens": 0.057344, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=12723410; data quality: config, params, base:Qwen/Qwen2.5-7B. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-7B" } ] }, { "id": "Qwen/Qwen2.5-1.5B-Instruct", "name": "Qwen2.5-1.5B-Instruct BF16", "short_name": "Qwen2.5 1.5B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 11679218, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "chat", "conversational", "en", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-1.5B", "base_model:finetune:Qwen/Qwen2.5-1.5B", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 1.5437, "active_params_b": 1.5437, "layers": 28, "attention_heads": 12, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 3.0874, "active_weight_gb": 3.0874, "kv_alloc_gb_per_1k_tokens": 0.028672, "kv_read_gb_per_1k_tokens": 0.028672, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=11679218; data quality: config, params, base:Qwen/Qwen2.5-1.5B. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B" } ] }, { "id": "google/gemma-4-31B-it", "name": "gemma-4-31B-it BF16", "short_name": "gemma 4 31B it", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 10921186, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "conversational", "base_model:google/gemma-4-31B", "base_model:finetune:google/gemma-4-31B", "license:apache-2.0", "eval-results", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 32.6824, "active_params_b": 32.6824, "layers": 60, "attention_heads": 32, "kv_heads": 16, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 65.3648, "active_weight_gb": 65.3648, "kv_alloc_gb_per_1k_tokens": 0.83968, "kv_read_gb_per_1k_tokens": 0.83968, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=10921186; data quality: config, params, base:google/gemma-4-31B. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-4-31B-it" }, { "label": "Model config", "url": "https://huggingface.co/google/gemma-4-31B-it/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-31B" } ] }, { "id": "meta-llama/Llama-3.1-8B-Instruct", "name": "Llama-3.1-8B-Instruct BF16", "short_name": "Llama 3.1 8B", "license": "llama3.1", "hf_pipeline_tag": "text-generation", "hf_downloads": 10008529, "tags": [ "transformers", "safetensors", "llama", "text-generation", "facebook", "meta", "pytorch", "llama-3", "conversational", "en", "de", "fr", "it", "pt", "hi", "es", "th", "arxiv:2204.05149", "base_model:meta-llama/Llama-3.1-8B", "base_model:finetune:meta-llama/Llama-3.1-8B", "license:llama3.1", "eval-results", "text-generation-inference", "endpoints_compatible" ], "architecture": { "type": "Dense transformer", "detail": "unknown", "total_params_b": 8.0303, "active_params_b": 8.0303, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": 2048, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 16.0606, "active_weight_gb": 16.0606, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=10008529; data quality: config, params, base:meta-llama/Llama-3.1-8B. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct" }, { "label": "Base model config", "url": "https://huggingface.co/meta-llama/Llama-3.1-8B/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/meta-llama/Llama-3.1-8B" } ] }, { "id": "Qwen/Qwen2.5-VL-7B-Instruct", "name": "Qwen2.5-VL-7B-Instruct BF16", "short_name": "Qwen2.5 VL 7B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 9705940, "tags": [ "transformers", "safetensors", "qwen2_5_vl", "image-text-to-text", "multimodal", "conversational", "en", "arxiv:2309.00071", "arxiv:2409.12191", "arxiv:2308.12966", "license:apache-2.0", "eval-results", "text-generation-inference", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen2_5_vl (Qwen2_5_VLForConditionalGeneration)", "total_params_b": 8.2922, "active_params_b": 8.2922, "layers": 28, "attention_heads": 28, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 128000, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 16.5844, "active_weight_gb": 16.5844, "kv_alloc_gb_per_1k_tokens": 0.057344, "kv_read_gb_per_1k_tokens": 0.057344, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=9705940; data quality: config, params. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct/raw/main/config.json" } ] }, { "id": "Qwen/Qwen3.5-9B", "name": "Qwen3.5-9B BF16/F32", "short_name": "Qwen3.5 9B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 8914120, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "conversational", "base_model:Qwen/Qwen3.5-9B-Base", "base_model:finetune:Qwen/Qwen3.5-9B-Base", "license:apache-2.0", "eval-results", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 9.6531, "active_params_b": 9.6531, "layers": 32, "attention_heads": 16, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16/F32", "weight_bytes_per_param": 2, "resident_weight_gb": 19.3062, "active_weight_gb": 19.3062, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=8914120; data quality: config, params, base:Qwen/Qwen3.5-9B-Base. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3.5-9B" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3.5-9B/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-9B-Base" } ] }, { "id": "meta-llama/Llama-3.2-1B-Instruct", "name": "Llama-3.2-1B-Instruct BF16", "short_name": "Llama 3.2 1B", "license": "llama3.2", "hf_pipeline_tag": "text-generation", "hf_downloads": 8583830, "tags": [ "transformers", "safetensors", "llama", "text-generation", "facebook", "meta", "pytorch", "llama-3", "conversational", "en", "de", "fr", "it", "pt", "hi", "es", "th", "arxiv:2204.05149", "arxiv:2405.16406", "license:llama3.2", "eval-results", "text-generation-inference", "endpoints_compatible" ], "architecture": { "type": "Dense transformer", "detail": "unknown", "total_params_b": 1.2358, "active_params_b": 1.2358, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": 2048, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 2.4716, "active_weight_gb": 2.4716, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=8583830; data quality: no-config, params. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct" } ] }, { "id": "Qwen/Qwen3.5-4B", "name": "Qwen3.5-4B BF16/F32", "short_name": "Qwen3.5 4B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 8395480, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "conversational", "base_model:Qwen/Qwen3.5-4B-Base", "base_model:finetune:Qwen/Qwen3.5-4B-Base", "license:apache-2.0", "eval-results", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 4.6599, "active_params_b": 4.6599, "layers": 32, "attention_heads": 16, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16/F32", "weight_bytes_per_param": 2, "resident_weight_gb": 9.3198, "active_weight_gb": 9.3198, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=8395480; data quality: config, params, base:Qwen/Qwen3.5-4B-Base. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3.5-4B" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3.5-4B/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-4B-Base" } ] }, { "id": "Qwen/Qwen2.5-3B-Instruct", "name": "Qwen2.5-3B-Instruct BF16", "short_name": "Qwen2.5 3B", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 7597019, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "chat", "conversational", "en", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-3B", "base_model:finetune:Qwen/Qwen2.5-3B", "license:other", "text-generation-inference", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 3.0859, "active_params_b": 3.0859, "layers": 36, "attention_heads": 16, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 6.1718, "active_weight_gb": 6.1718, "kv_alloc_gb_per_1k_tokens": 0.036864, "kv_read_gb_per_1k_tokens": 0.036864, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=7597019; data quality: config, params, base:Qwen/Qwen2.5-3B. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-3B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-3B-Instruct/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-3B" } ] }, { "id": "deepseek-ai/DeepSeek-R1", "name": "DeepSeek-R1 FP8", "short_name": "DeepSeek R1", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 7483926, "tags": [ "transformers", "safetensors", "deepseek_v3", "text-generation", "conversational", "custom_code", "arxiv:2501.12948", "license:mit", "eval-results", "text-generation-inference", "endpoints_compatible", "fp8" ], "architecture": { "type": "MoE transformer", "detail": "deepseek_v3 (DeepseekV3ForCausalLM)", "total_params_b": 684.5314, "active_params_b": 21.3916, "layers": 61, "attention_heads": 128, "kv_heads": 128, "head_dim": 56, "max_context_tokens": 163840, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "FP8", "weight_bytes_per_param": 1, "resident_weight_gb": 684.5314, "active_weight_gb": 21.3916, "kv_alloc_gb_per_1k_tokens": 1.748992, "kv_read_gb_per_1k_tokens": 1.748992, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=7483926; data quality: config, params. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1" }, { "label": "Model config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1/raw/main/config.json" } ] }, { "id": "openai/gpt-oss-20b", "name": "gpt-oss-20b MXFP4", "short_name": "gpt oss 20b", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 6973137, "tags": [ "transformers", "safetensors", "gpt_oss", "text-generation", "vllm", "conversational", "arxiv:2508.10925", "license:apache-2.0", "eval-results", "endpoints_compatible", "8-bit", "mxfp4", "deploy:azure" ], "architecture": { "type": "MoE transformer", "detail": "gpt_oss (GptOssForCausalLM)", "total_params_b": 21.512, "active_params_b": 2.689, "layers": 24, "attention_heads": 64, "kv_heads": 8, "head_dim": 64, "max_context_tokens": 131072, "routed_experts": 32, "routed_experts_per_token": 4, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "MXFP4", "weight_bytes_per_param": 0.5, "resident_weight_gb": 10.756, "active_weight_gb": 1.3445, "kv_alloc_gb_per_1k_tokens": 0.049152, "kv_read_gb_per_1k_tokens": 0.049152, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=6973137; data quality: config, params. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/openai/gpt-oss-20b" }, { "label": "Model config", "url": "https://huggingface.co/openai/gpt-oss-20b/raw/main/config.json" } ] }, { "id": "antirez/deepseek-v4-gguf", "name": "deepseek-v4-gguf GGUF 2-bit", "short_name": "deepseek v4", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 6441753, "tags": [ "gguf", "quantized", "deepseek", "deepseek-v4", "deepseek-v4-flash", "moe", "mixture-of-experts", "2-bit", "4-bit", "iq2_xxs", "q2_k", "q4_k", "ds4", "apple-silicon", "metal", "text-generation", "en", "base_model:deepseek-ai/DeepSeek-V4-Flash", "base_model:quantized:deepseek-ai/DeepSeek-V4-Flash", "license:mit", "endpoints_compatible", "conversational" ], "architecture": { "type": "MoE transformer", "detail": "deepseek_v4 (DeepseekV4ForCausalLM)", "total_params_b": 158.0694, "active_params_b": 3.7048, "layers": 43, "attention_heads": 64, "kv_heads": 1, "head_dim": 512, "max_context_tokens": 1048576, "routed_experts": 256, "routed_experts_per_token": 6, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "GGUF 2-bit", "weight_bytes_per_param": 0.25, "resident_weight_gb": 39.5174, "active_weight_gb": 0.9262, "kv_alloc_gb_per_1k_tokens": 0.088064, "kv_read_gb_per_1k_tokens": 0.088064, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=6441753; data quality: config, params, base:deepseek-ai/DeepSeek-V4-Flash. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/antirez/deepseek-v4-gguf" }, { "label": "Base model config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash" } ] }, { "id": "google/gemma-3-270m", "name": "gemma-3-270m BF16", "short_name": "gemma 3 270m", "license": "gemma", "hf_pipeline_tag": "text-generation", "hf_downloads": 5982746, "tags": [ "transformers", "safetensors", "gemma3", "gemma", "google", "text-generation", "arxiv:2503.19786", "arxiv:1905.07830", "arxiv:1905.10044", "arxiv:1911.11641", "arxiv:1705.03551", "arxiv:1911.01547", "arxiv:1907.10641", "arxiv:2311.07911", "arxiv:2311.12022", "arxiv:2411.04368", "arxiv:1904.09728", "arxiv:1903.00161", "arxiv:2009.03300", "arxiv:2304.06364", "arxiv:2103.03874", "arxiv:2110.14168", "arxiv:2108.07732", "arxiv:2107.03374", "arxiv:2403.07974", "arxiv:2305.03111", "arxiv:2405.04520", "arxiv:2210.03057", "arxiv:2106.03193", "arxiv:1910.11856", "arxiv:2502.12404", "arxiv:2502.21228", "arxiv:2404.16816", "arxiv:2104.12756", "arxiv:2311.16502", "arxiv:2203.10244", "arxiv:2404.12390", "arxiv:1810.12440", "arxiv:1908.02660", "arxiv:2310.02255", "arxiv:2312.11805", "license:gemma", "endpoints_compatible" ], "architecture": { "type": "Dense transformer", "detail": "unknown", "total_params_b": 0.2681, "active_params_b": 0.2681, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": 2048, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 0.5362, "active_weight_gb": 0.5362, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=5982746; data quality: no-config, params. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-3-270m" } ] }, { "id": "Qwen/Qwen3.6-35B-A3B-FP8", "name": "Qwen3.6-35B-A3B-FP8 FP8", "short_name": "Qwen3.6 35B A3B FP8", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 5952150, "tags": [ "transformers", "safetensors", "qwen3_5_moe", "image-text-to-text", "conversational", "base_model:Qwen/Qwen3.6-35B-A3B", "base_model:quantized:Qwen/Qwen3.6-35B-A3B", "license:apache-2.0", "endpoints_compatible", "fp8", "deploy:azure" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_5_moe_text", "total_params_b": 35.9539, "active_params_b": 3, "layers": 40, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "FP8", "weight_bytes_per_param": 1, "resident_weight_gb": 35.9539, "active_weight_gb": 3, "kv_alloc_gb_per_1k_tokens": 0.08192, "kv_read_gb_per_1k_tokens": 0.08192, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=5952150; data quality: config, params, base:Qwen/Qwen3.6-35B-A3B. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B-FP8" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B-FP8/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B" } ] }, { "id": "Qwen/Qwen3-1.7B", "name": "Qwen3-1.7B BF16", "short_name": "Qwen3 1.7B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 5698326, "tags": [ "transformers", "safetensors", "qwen3", "text-generation", "conversational", "arxiv:2505.09388", "base_model:Qwen/Qwen3-1.7B-Base", "base_model:finetune:Qwen/Qwen3-1.7B-Base", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 2.0317, "active_params_b": 2.0317, "layers": 28, "attention_heads": 16, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 4.0634, "active_weight_gb": 4.0634, "kv_alloc_gb_per_1k_tokens": 0.114688, "kv_read_gb_per_1k_tokens": 0.114688, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=5698326; data quality: config, params, base:Qwen/Qwen3-1.7B-Base. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-1.7B" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-1.7B/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-1.7B-Base" } ] }, { "id": "Qwen/Qwen3.6-35B-A3B", "name": "Qwen3.6-35B-A3B BF16", "short_name": "Qwen3.6 35B A3B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 5656573, "tags": [ "transformers", "safetensors", "qwen3_5_moe", "image-text-to-text", "conversational", "license:apache-2.0", "eval-results", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_5_moe_text", "total_params_b": 35.9518, "active_params_b": 3, "layers": 40, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 71.9036, "active_weight_gb": 6, "kv_alloc_gb_per_1k_tokens": 0.08192, "kv_read_gb_per_1k_tokens": 0.08192, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=5656573; data quality: config, params. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/main/config.json" } ] }, { "id": "google/gemma-4-E4B-it", "name": "gemma-4-E4B-it BF16", "short_name": "gemma 4 E4B it", "license": "apache-2.0", "hf_pipeline_tag": "any-to-any", "hf_downloads": 5541055, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "any-to-any", "base_model:google/gemma-4-E4B", "base_model:finetune:google/gemma-4-E4B", "license:apache-2.0", "eval-results", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 7.9962, "active_params_b": 7.9962, "layers": 42, "attention_heads": 8, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 15.9924, "active_weight_gb": 15.9924, "kv_alloc_gb_per_1k_tokens": 0.086016, "kv_read_gb_per_1k_tokens": 0.086016, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=5541055; data quality: config, params, base:google/gemma-4-E4B. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-4-E4B-it" }, { "label": "Model config", "url": "https://huggingface.co/google/gemma-4-E4B-it/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-E4B" } ] }, { "id": "nvidia/Qwen3.6-35B-A3B-NVFP4", "name": "Qwen3.6-35B-A3B-NVFP4 NVFP4", "short_name": "Qwen3.6 35B A3B NVFP4", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 5495402, "tags": [ "Model Optimizer", "safetensors", "qwen3_5_moe", "nvidia", "ModelOpt", "Qwen3.6", "quantized", "FP4", "fp4", "text-generation", "conversational", "base_model:Qwen/Qwen3.6-35B-A3B", "base_model:quantized:Qwen/Qwen3.6-35B-A3B", "license:apache-2.0", "8-bit", "modelopt", "deploy:azure" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_5_moe_text", "total_params_b": 35.9518, "active_params_b": 3, "layers": 40, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "NVFP4", "weight_bytes_per_param": 0.5, "resident_weight_gb": 17.9759, "active_weight_gb": 1.5, "kv_alloc_gb_per_1k_tokens": 0.08192, "kv_read_gb_per_1k_tokens": 0.08192, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=5495402; data quality: config, params, base:Qwen/Qwen3.6-35B-A3B. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/Qwen3.6-35B-A3B-NVFP4" }, { "label": "Model config", "url": "https://huggingface.co/nvidia/Qwen3.6-35B-A3B-NVFP4/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B" } ] }, { "id": "Qwen/Qwen3-4B-Instruct-2507", "name": "Qwen3-4B-Instruct-2507 BF16", "short_name": "Qwen3 4B Instruct 2507", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 5299649, "tags": [ "transformers", "safetensors", "qwen3", "text-generation", "conversational", "arxiv:2505.09388", "license:apache-2.0", "eval-results", "text-generation-inference", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 4.0225, "active_params_b": 4.0225, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 8.045, "active_weight_gb": 8.045, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=5299649; data quality: config, params. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507/raw/main/config.json" } ] }, { "id": "Qwen/Qwen3.6-27B", "name": "Qwen3.6-27B BF16", "short_name": "Qwen3.6 27B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 5260258, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "conversational", "license:apache-2.0", "eval-results", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 27.7814, "active_params_b": 27.7814, "layers": 64, "attention_heads": 24, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 55.5628, "active_weight_gb": 55.5628, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=5260258; data quality: config, params. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3.6-27B" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/main/config.json" } ] }, { "id": "Qwen/Qwen2.5-VL-3B-Instruct", "name": "Qwen2.5-VL-3B-Instruct BF16", "short_name": "Qwen2.5 VL 3B", "license": "unknown", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 5237979, "tags": [ "transformers", "safetensors", "qwen2_5_vl", "image-text-to-text", "multimodal", "conversational", "en", "arxiv:2309.00071", "arxiv:2409.12191", "arxiv:2308.12966", "eval-results", "text-generation-inference", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen2_5_vl (Qwen2_5_VLForConditionalGeneration)", "total_params_b": 3.7546, "active_params_b": 3.7546, "layers": 36, "attention_heads": 16, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 128000, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 7.5092, "active_weight_gb": 7.5092, "kv_alloc_gb_per_1k_tokens": 0.036864, "kv_read_gb_per_1k_tokens": 0.036864, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=5237979; data quality: config, params. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct/raw/main/config.json" } ] }, { "id": "Qwen/Qwen3-VL-8B-Instruct", "name": "Qwen3-VL-8B-Instruct BF16", "short_name": "Qwen3 VL 8B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 5180106, "tags": [ "transformers", "safetensors", "qwen3_vl", "image-text-to-text", "conversational", "arxiv:2505.09388", "arxiv:2502.13923", "arxiv:2409.12191", "arxiv:2308.12966", "license:apache-2.0", "eval-results", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_vl_text", "total_params_b": 8.7671, "active_params_b": 8.7671, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 17.5342, "active_weight_gb": 17.5342, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=5180106; data quality: config, params. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct/raw/main/config.json" } ] }, { "id": "Qwen/Qwen3-VL-32B-Instruct", "name": "Qwen3-VL-32B-Instruct BF16", "short_name": "Qwen3 VL 32B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 5081597, "tags": [ "transformers", "safetensors", "qwen3_vl", "image-text-to-text", "conversational", "arxiv:2505.09388", "arxiv:2502.13923", "arxiv:2409.12191", "arxiv:2308.12966", "license:apache-2.0", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_vl_text", "total_params_b": 33.3574, "active_params_b": 33.3574, "layers": 64, "attention_heads": 64, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 66.7148, "active_weight_gb": 66.7148, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=5081597; data quality: config, params. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-32B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-VL-32B-Instruct/raw/main/config.json" } ] }, { "id": "Qwen/Qwen2.5-0.5B-Instruct", "name": "Qwen2.5-0.5B-Instruct BF16", "short_name": "Qwen2.5 0.5B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 5008910, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "chat", "conversational", "en", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-0.5B", "base_model:finetune:Qwen/Qwen2.5-0.5B", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 0.494, "active_params_b": 0.494, "layers": 24, "attention_heads": 14, "kv_heads": 2, "head_dim": 64, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 0.988, "active_weight_gb": 0.988, "kv_alloc_gb_per_1k_tokens": 0.012288, "kv_read_gb_per_1k_tokens": 0.012288, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=5008910; data quality: config, params, base:Qwen/Qwen2.5-0.5B. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-0.5B" } ] }, { "id": "Qwen/Qwen3.6-27B-FP8", "name": "Qwen3.6-27B-FP8 FP8", "short_name": "Qwen3.6 27B FP8", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 4821147, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "conversational", "base_model:Qwen/Qwen3.6-27B", "base_model:quantized:Qwen/Qwen3.6-27B", "license:apache-2.0", "endpoints_compatible", "fp8", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 27.7829, "active_params_b": 27.7829, "layers": 64, "attention_heads": 24, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "FP8", "weight_bytes_per_param": 1, "resident_weight_gb": 27.7829, "active_weight_gb": 27.7829, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=4821147; data quality: config, params, base:Qwen/Qwen3.6-27B. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3.6-27B-FP8" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3.6-27B-FP8/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.6-27B" } ] }, { "id": "cyankiwi/gemma-4-26B-A4B-it-AWQ-4bit", "name": "gemma-4-26B-A4B-it-AWQ-4bit 4-bit", "short_name": "gemma 4 26B A4B it AWQ 4bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 4707741, "tags": [ "transformers", "safetensors", "image-text-to-text", "base_model:google/gemma-4-26B-A4B-it", "base_model:finetune:google/gemma-4-26B-A4B-it", "license:apache-2.0", "endpoints_compatible" ], "architecture": { "type": "MoE transformer", "detail": "gemma4_text", "total_params_b": 26.5543, "active_params_b": 4, "layers": 30, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 1, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "4-bit", "weight_bytes_per_param": 0.5, "resident_weight_gb": 13.2772, "active_weight_gb": 2, "kv_alloc_gb_per_1k_tokens": 0.20992, "kv_read_gb_per_1k_tokens": 0.20992, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=4707741; data quality: config, params, base:google/gemma-4-26B-A4B-it. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/cyankiwi/gemma-4-26B-A4B-it-AWQ-4bit" }, { "label": "Model config", "url": "https://huggingface.co/cyankiwi/gemma-4-26B-A4B-it-AWQ-4bit/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it" } ] }, { "id": "Qwen/Qwen3-32B", "name": "Qwen3-32B BF16", "short_name": "Qwen3 32B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 4276519, "tags": [ "transformers", "safetensors", "qwen3", "text-generation", "conversational", "arxiv:2309.00071", "arxiv:2505.09388", "license:apache-2.0", "eval-results", "text-generation-inference", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 32.7621, "active_params_b": 32.7621, "layers": 64, "attention_heads": 64, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 65.5242, "active_weight_gb": 65.5242, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=4276519; data quality: config, params. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-32B" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-32B/raw/main/config.json" } ] }, { "id": "Qwen/Qwen2.5-Coder-14B-Instruct", "name": "Qwen2.5-Coder-14B-Instruct BF16", "short_name": "Qwen2.5 Coder 14B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 4155102, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "code", "codeqwen", "chat", "qwen", "qwen-coder", "conversational", "en", "arxiv:2409.12186", "arxiv:2309.00071", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-Coder-14B", "base_model:finetune:Qwen/Qwen2.5-Coder-14B", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 14.77, "active_params_b": 14.77, "layers": 48, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 29.54, "active_weight_gb": 29.54, "kv_alloc_gb_per_1k_tokens": 0.196608, "kv_read_gb_per_1k_tokens": 0.196608, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=4155102; data quality: config, params, base:Qwen/Qwen2.5-Coder-14B. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-14B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-14B-Instruct/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-14B" } ] }, { "id": "openai/gpt-oss-120b", "name": "gpt-oss-120b MXFP4", "short_name": "gpt oss 120b", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 4020164, "tags": [ "transformers", "safetensors", "gpt_oss", "text-generation", "vllm", "conversational", "arxiv:2508.10925", "license:apache-2.0", "eval-results", "endpoints_compatible", "8-bit", "mxfp4", "deploy:azure" ], "architecture": { "type": "MoE transformer", "detail": "gpt_oss (GptOssForCausalLM)", "total_params_b": 120.4123, "active_params_b": 3.7629, "layers": 36, "attention_heads": 64, "kv_heads": 8, "head_dim": 64, "max_context_tokens": 131072, "routed_experts": 128, "routed_experts_per_token": 4, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "MXFP4", "weight_bytes_per_param": 0.5, "resident_weight_gb": 60.2062, "active_weight_gb": 1.8815, "kv_alloc_gb_per_1k_tokens": 0.073728, "kv_read_gb_per_1k_tokens": 0.073728, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=4020164; data quality: config, params. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/openai/gpt-oss-120b" }, { "label": "Model config", "url": "https://huggingface.co/openai/gpt-oss-120b/raw/main/config.json" } ] }, { "id": "google/gemma-3-1b-it", "name": "gemma-3-1b-it BF16", "short_name": "gemma 3 1b it", "license": "gemma", "hf_pipeline_tag": "text-generation", "hf_downloads": 4017683, "tags": [ "transformers", "safetensors", "gemma3_text", "text-generation", "conversational", "arxiv:1905.07830", "arxiv:1905.10044", "arxiv:1911.11641", "arxiv:1904.09728", "arxiv:1705.03551", "arxiv:1911.01547", "arxiv:1907.10641", "arxiv:1903.00161", "arxiv:2009.03300", "arxiv:2304.06364", "arxiv:2103.03874", "arxiv:2110.14168", "arxiv:2311.12022", "arxiv:2108.07732", "arxiv:2107.03374", "arxiv:2210.03057", "arxiv:2106.03193", "arxiv:1910.11856", "arxiv:2502.12404", "arxiv:2502.21228", "arxiv:2404.16816", "arxiv:2104.12756", "arxiv:2311.16502", "arxiv:2203.10244", "arxiv:2404.12390", "arxiv:1810.12440", "arxiv:1908.02660", "arxiv:2312.11805", "base_model:google/gemma-3-1b-pt", "base_model:finetune:google/gemma-3-1b-pt", "license:gemma", "eval-results", "text-generation-inference", "endpoints_compatible" ], "architecture": { "type": "Dense transformer", "detail": "unknown", "total_params_b": 0.9999, "active_params_b": 0.9999, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": 2048, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 1.9998, "active_weight_gb": 1.9998, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=4017683; data quality: config, params, base:google/gemma-3-1b-pt. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-3-1b-it" }, { "label": "Base model config", "url": "https://huggingface.co/google/gemma-3-1b-pt/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-3-1b-pt" } ] }, { "id": "Qwen/Qwen3-VL-4B-Instruct", "name": "Qwen3-VL-4B-Instruct BF16", "short_name": "Qwen3 VL 4B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 3980238, "tags": [ "transformers", "safetensors", "qwen3_vl", "image-text-to-text", "conversational", "arxiv:2505.09388", "arxiv:2502.13923", "arxiv:2409.12191", "arxiv:2308.12966", "license:apache-2.0", "eval-results", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_vl_text", "total_params_b": 4.4378, "active_params_b": 4.4378, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 8.8756, "active_weight_gb": 8.8756, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=3980238; data quality: config, params. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct/raw/main/config.json" } ] }, { "id": "Qwen/Qwen2-1.5B-Instruct", "name": "Qwen2-1.5B-Instruct BF16", "short_name": "Qwen2 1.5B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 3730070, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "chat", "conversational", "en", "license:apache-2.0", "eval-results", "text-generation-inference", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 1.5437, "active_params_b": 1.5437, "layers": 28, "attention_heads": 12, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 3.0874, "active_weight_gb": 3.0874, "kv_alloc_gb_per_1k_tokens": 0.028672, "kv_read_gb_per_1k_tokens": 0.028672, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=3730070; data quality: config, params. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2-1.5B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2-1.5B-Instruct/raw/main/config.json" } ] }, { "id": "Qwen/Qwen2-VL-2B-Instruct", "name": "Qwen2-VL-2B-Instruct BF16", "short_name": "Qwen2 VL 2B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 3675633, "tags": [ "transformers", "safetensors", "qwen2_vl", "image-text-to-text", "multimodal", "conversational", "en", "arxiv:2409.12191", "arxiv:2308.12966", "base_model:Qwen/Qwen2-VL-2B", "base_model:finetune:Qwen/Qwen2-VL-2B", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen2_vl (Qwen2VLForConditionalGeneration)", "total_params_b": 2.209, "active_params_b": 2.209, "layers": 28, "attention_heads": 12, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 4.418, "active_weight_gb": 4.418, "kv_alloc_gb_per_1k_tokens": 0.028672, "kv_read_gb_per_1k_tokens": 0.028672, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=3675633; data quality: config, params, base:Qwen/Qwen2-VL-2B. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2-VL-2B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2-VL-2B-Instruct/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2-VL-2B" } ] }, { "id": "Qwen/Qwen3-14B", "name": "Qwen3-14B BF16", "short_name": "Qwen3 14B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 3651617, "tags": [ "transformers", "safetensors", "qwen3", "text-generation", "conversational", "arxiv:2309.00071", "arxiv:2505.09388", "base_model:Qwen/Qwen3-14B-Base", "base_model:finetune:Qwen/Qwen3-14B-Base", "license:apache-2.0", "text-generation-inference", "endpoints_compatible" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 14.7683, "active_params_b": 14.7683, "layers": 40, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 29.5366, "active_weight_gb": 29.5366, "kv_alloc_gb_per_1k_tokens": 0.16384, "kv_read_gb_per_1k_tokens": 0.16384, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=3651617; data quality: config, params, base:Qwen/Qwen3-14B-Base. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-14B" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-14B/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-14B-Base" } ] }, { "id": "deepseek-ai/DeepSeek-R1-0528", "name": "DeepSeek-R1-0528 FP8", "short_name": "DeepSeek R1 0528", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 3586665, "tags": [ "transformers", "safetensors", "deepseek_v3", "text-generation", "conversational", "custom_code", "arxiv:2501.12948", "license:mit", "eval-results", "text-generation-inference", "endpoints_compatible", "fp8" ], "architecture": { "type": "MoE transformer", "detail": "deepseek_v3 (DeepseekV3ForCausalLM)", "total_params_b": 684.5314, "active_params_b": 21.3916, "layers": 61, "attention_heads": 128, "kv_heads": 128, "head_dim": 56, "max_context_tokens": 163840, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "FP8", "weight_bytes_per_param": 1, "resident_weight_gb": 684.5314, "active_weight_gb": 21.3916, "kv_alloc_gb_per_1k_tokens": 1.748992, "kv_read_gb_per_1k_tokens": 1.748992, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=3586665; data quality: config, params. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-0528" }, { "label": "Model config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/config.json" } ] }, { "id": "Qwen/Qwen2.5-7B-Instruct-AWQ", "name": "Qwen2.5-7B-Instruct-AWQ 4-bit", "short_name": "Qwen2.5 7B Instruct", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 3476715, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "chat", "conversational", "en", "arxiv:2309.00071", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-7B-Instruct", "base_model:quantized:Qwen/Qwen2.5-7B-Instruct", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "4-bit", "awq" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 7.6156, "active_params_b": 7.6156, "layers": 28, "attention_heads": 28, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "4-bit", "weight_bytes_per_param": 0.5, "resident_weight_gb": 3.8078, "active_weight_gb": 3.8078, "kv_alloc_gb_per_1k_tokens": 0.057344, "kv_read_gb_per_1k_tokens": 0.057344, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=3476715; data quality: config, params, base:Qwen/Qwen2.5-7B-Instruct. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct-AWQ" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct-AWQ/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct" } ] }, { "id": "Andycurrent/Gemma-3-1B-it-GLM-4.7-Flash-Heretic-Uncensored-Thinking_GGUF", "name": "Gemma-3-1B-it-GLM-4.7-Flash-Heretic-Uncensored-Thinking_GGUF GGUF", "short_name": "Gemma 3 1B it GLM 4.7 Flash Heretic Uncensored T", "license": "gemma", "hf_pipeline_tag": "text-generation", "hf_downloads": 3080875, "tags": [ "gguf", "uncensored", "text-generation", "reasoning", "instruction-tuned", "lightweight", "en", "base_model:google/gemma-3-1b-it", "base_model:quantized:google/gemma-3-1b-it", "license:gemma", "endpoints_compatible", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "unknown", "total_params_b": 0.9999, "active_params_b": 0.9999, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": 2048, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF", "weight_bytes_per_param": 2, "resident_weight_gb": 1.9998, "active_weight_gb": 1.9998, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=3080875; data quality: config, params, base:google/gemma-3-1b-it. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Andycurrent/Gemma-3-1B-it-GLM-4.7-Flash-Heretic-Uncensored-Thinking_GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/google/gemma-3-1b-it/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-3-1b-it" } ] }, { "id": "Qwen/Qwen2.5-32B-Instruct", "name": "Qwen2.5-32B-Instruct BF16", "short_name": "Qwen2.5 32B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 2946692, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "chat", "conversational", "en", "arxiv:2309.00071", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-32B", "base_model:finetune:Qwen/Qwen2.5-32B", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 32.7639, "active_params_b": 32.7639, "layers": 64, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 65.5278, "active_weight_gb": 65.5278, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=2946692; data quality: config, params, base:Qwen/Qwen2.5-32B. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-32B" } ] }, { "id": "Qwen/Qwen3-30B-A3B", "name": "Qwen3-30B-A3B BF16", "short_name": "Qwen3 30B A3B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 2808933, "tags": [ "transformers", "safetensors", "text-generation", "arxiv:2309.00071", "arxiv:2505.09388", "base_model:Qwen/Qwen3-30B-A3B-Base", "base_model:finetune:Qwen/Qwen3-30B-A3B-Base", "license:apache-2.0", "endpoints_compatible" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_moe (Qwen3MoeForCausalLM)", "total_params_b": 30.5321, "active_params_b": 3, "layers": 48, "attention_heads": 32, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 61.0642, "active_weight_gb": 6, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=2808933; data quality: config, params, base:Qwen/Qwen3-30B-A3B-Base. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Base" } ] }, { "id": "google/gemma-4-12B-it", "name": "gemma-4-12B-it BF16", "short_name": "gemma 4 12B it", "license": "apache-2.0", "hf_pipeline_tag": "any-to-any", "hf_downloads": 2621462, "tags": [ "transformers", "safetensors", "gemma4_unified", "image-text-to-text", "any-to-any", "base_model:google/gemma-4-12B", "base_model:finetune:google/gemma-4-12B", "license:apache-2.0", "eval-results", "endpoints_compatible" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_unified_text", "total_params_b": 11.9597, "active_params_b": 11.9597, "layers": 48, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 23.9194, "active_weight_gb": 23.9194, "kv_alloc_gb_per_1k_tokens": 0.393216, "kv_read_gb_per_1k_tokens": 0.393216, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=2621462; data quality: config, params, base:google/gemma-4-12B. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-4-12B-it" }, { "label": "Model config", "url": "https://huggingface.co/google/gemma-4-12B-it/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-12B" } ] }, { "id": "microsoft/Florence-2-base", "name": "Florence-2-base F16", "short_name": "Florence 2", "license": "mit", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 2590016, "tags": [ "transformers", "pytorch", "safetensors", "florence2", "image-text-to-text", "vision", "custom_code", "arxiv:2311.06242", "license:mit", "endpoints_compatible" ], "architecture": { "type": "Dense transformer", "detail": "unknown", "total_params_b": 0.2316, "active_params_b": 0.2316, "layers": 6, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": 1024, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "F16", "weight_bytes_per_param": 2, "resident_weight_gb": 0.4632, "active_weight_gb": 0.4632, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=2590016; data quality: config, params. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/microsoft/Florence-2-base" }, { "label": "Model config", "url": "https://huggingface.co/microsoft/Florence-2-base/raw/main/config.json" } ] }, { "id": "Qwen/Qwen3.5-27B", "name": "Qwen3.5-27B BF16/F32", "short_name": "Qwen3.5 27B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 2560332, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "conversational", "license:apache-2.0", "eval-results", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 27.7814, "active_params_b": 27.7814, "layers": 64, "attention_heads": 24, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16/F32", "weight_bytes_per_param": 2, "resident_weight_gb": 55.5628, "active_weight_gb": 55.5628, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=2560332; data quality: config, params. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3.5-27B" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3.5-27B/raw/main/config.json" } ] }, { "id": "Qwen/Qwen3.5-0.8B", "name": "Qwen3.5-0.8B F32/BF16", "short_name": "Qwen3.5 0.8B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 2510164, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "conversational", "base_model:Qwen/Qwen3.5-0.8B-Base", "base_model:finetune:Qwen/Qwen3.5-0.8B-Base", "license:apache-2.0", "eval-results", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 0.8734, "active_params_b": 0.8734, "layers": 24, "attention_heads": 8, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "F32/BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 1.7468, "active_weight_gb": 1.7468, "kv_alloc_gb_per_1k_tokens": 0.049152, "kv_read_gb_per_1k_tokens": 0.049152, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=2510164; data quality: config, params, base:Qwen/Qwen3.5-0.8B-Base. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3.5-0.8B" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3.5-0.8B/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-0.8B-Base" } ] }, { "id": "deepseek-ai/DeepSeek-V3.2", "name": "DeepSeek-V3.2 FP8", "short_name": "DeepSeek V3.2", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 2391599, "tags": [ "transformers", "safetensors", "deepseek_v32", "text-generation", "conversational", "base_model:deepseek-ai/DeepSeek-V3.2-Exp-Base", "base_model:finetune:deepseek-ai/DeepSeek-V3.2-Exp-Base", "license:mit", "eval-results", "endpoints_compatible", "fp8" ], "architecture": { "type": "MoE transformer", "detail": "deepseek_v32 (DeepseekV32ForCausalLM)", "total_params_b": 685.3969, "active_params_b": 21.4187, "layers": 61, "attention_heads": 128, "kv_heads": 128, "head_dim": 56, "max_context_tokens": 163840, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "FP8", "weight_bytes_per_param": 1, "resident_weight_gb": 685.3969, "active_weight_gb": 21.4187, "kv_alloc_gb_per_1k_tokens": 1.748992, "kv_read_gb_per_1k_tokens": 1.748992, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=2391599; data quality: config, params, base:deepseek-ai/DeepSeek-V3.2-Exp-Base. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.2" }, { "label": "Model config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.2/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp-Base" } ] }, { "id": "RedHatAI/gemma-4-31B-it-FP8-block", "name": "gemma-4-31B-it-FP8-block FP8", "short_name": "gemma 4 31B it FP8 block", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 2374319, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "fp8", "vllm", "llm-compressor", "compressed-tensors", "conversational", "base_model:google/gemma-4-31B-it", "base_model:quantized:google/gemma-4-31B-it", "license:apache-2.0", "endpoints_compatible" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 32.6824, "active_params_b": 32.6824, "layers": 60, "attention_heads": 32, "kv_heads": 16, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "FP8", "weight_bytes_per_param": 1, "resident_weight_gb": 32.6824, "active_weight_gb": 32.6824, "kv_alloc_gb_per_1k_tokens": 0.83968, "kv_read_gb_per_1k_tokens": 0.83968, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=2374319; data quality: config, params, base:google/gemma-4-31B-it. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/RedHatAI/gemma-4-31B-it-FP8-block" }, { "label": "Model config", "url": "https://huggingface.co/RedHatAI/gemma-4-31B-it-FP8-block/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-31B-it" } ] }, { "id": "zai-org/GLM-4.7-Flash", "name": "GLM-4.7-Flash BF16/F32", "short_name": "GLM 4.7 Flash", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 2354280, "tags": [ "transformers", "safetensors", "glm4_moe_lite", "text-generation", "conversational", "en", "zh", "arxiv:2508.06471", "license:mit", "eval-results", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "MoE transformer", "detail": "glm4_moe_lite (Glm4MoeLiteForCausalLM)", "total_params_b": 31.2215, "active_params_b": 1.9513, "layers": 47, "attention_heads": 20, "kv_heads": 20, "head_dim": 102, "max_context_tokens": 202752, "routed_experts": 64, "routed_experts_per_token": 4, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "BF16/F32", "weight_bytes_per_param": 2, "resident_weight_gb": 62.443, "active_weight_gb": 3.9026, "kv_alloc_gb_per_1k_tokens": 0.38352, "kv_read_gb_per_1k_tokens": 0.38352, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=2354280; data quality: config, params. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/zai-org/GLM-4.7-Flash" }, { "label": "Model config", "url": "https://huggingface.co/zai-org/GLM-4.7-Flash/raw/main/config.json" } ] }, { "id": "moonshotai/Kimi-K2.6", "name": "Kimi-K2.6 4-bit", "short_name": "Kimi K2.6", "license": "other", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 2302910, "tags": [ "transformers", "safetensors", "kimi_k25", "feature-extraction", "compressed-tensors", "image-text-to-text", "conversational", "custom_code", "arxiv:2602.02276", "license:other", "eval-results" ], "architecture": { "type": "MoE transformer", "detail": "kimi_k2 (DeepseekV3ForCausalLM)", "total_params_b": 1058.5894, "active_params_b": 22.0539, "layers": 61, "attention_heads": 64, "kv_heads": 64, "head_dim": 112, "max_context_tokens": 262144, "routed_experts": 384, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "4-bit", "weight_bytes_per_param": 0.5, "resident_weight_gb": 529.2947, "active_weight_gb": 11.0269, "kv_alloc_gb_per_1k_tokens": 1.748992, "kv_read_gb_per_1k_tokens": 1.748992, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=2302910; data quality: config, params. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/moonshotai/Kimi-K2.6" }, { "label": "Model config", "url": "https://huggingface.co/moonshotai/Kimi-K2.6/raw/main/config.json" } ] }, { "id": "google/gemma-4-E2B-it", "name": "gemma-4-E2B-it BF16", "short_name": "gemma 4 E2B it", "license": "apache-2.0", "hf_pipeline_tag": "any-to-any", "hf_downloads": 2285043, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "any-to-any", "base_model:google/gemma-4-E2B", "base_model:finetune:google/gemma-4-E2B", "license:apache-2.0", "eval-results", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 5.1232, "active_params_b": 5.1232, "layers": 35, "attention_heads": 8, "kv_heads": 1, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 10.2464, "active_weight_gb": 10.2464, "kv_alloc_gb_per_1k_tokens": 0.03584, "kv_read_gb_per_1k_tokens": 0.03584, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=2285043; data quality: config, params, base:google/gemma-4-E2B. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-4-E2B-it" }, { "label": "Model config", "url": "https://huggingface.co/google/gemma-4-E2B-it/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-E2B" } ] }, { "id": "zai-org/GLM-5-FP8", "name": "GLM-5-FP8 FP8", "short_name": "GLM 5 FP8", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 2283398, "tags": [ "transformers", "safetensors", "glm_moe_dsa", "text-generation", "conversational", "en", "zh", "arxiv:2602.15763", "license:mit", "eval-results", "endpoints_compatible", "fp8" ], "architecture": { "type": "MoE transformer", "detail": "glm_moe_dsa (GlmMoeDsaForCausalLM)", "total_params_b": 753.91, "active_params_b": 23.5597, "layers": 78, "attention_heads": 64, "kv_heads": 64, "head_dim": 64, "max_context_tokens": 202752, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "FP8", "weight_bytes_per_param": 1, "resident_weight_gb": 753.91, "active_weight_gb": 23.5597, "kv_alloc_gb_per_1k_tokens": 1.277952, "kv_read_gb_per_1k_tokens": 1.277952, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=2283398; data quality: config, params. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/zai-org/GLM-5-FP8" }, { "label": "Model config", "url": "https://huggingface.co/zai-org/GLM-5-FP8/raw/main/config.json" } ] }, { "id": "Qwen/Qwen2.5-0.5B", "name": "Qwen2.5-0.5B BF16", "short_name": "Qwen2.5 0.5B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 2280458, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "conversational", "en", "arxiv:2407.10671", "license:apache-2.0", "eval-results", "text-generation-inference", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 0.494, "active_params_b": 0.494, "layers": 24, "attention_heads": 14, "kv_heads": 2, "head_dim": 64, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 0.988, "active_weight_gb": 0.988, "kv_alloc_gb_per_1k_tokens": 0.012288, "kv_read_gb_per_1k_tokens": 0.012288, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=2280458; data quality: config, params. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-0.5B" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-0.5B/raw/main/config.json" } ] }, { "id": "RedHatAI/Qwen3.6-35B-A3B-NVFP4", "name": "Qwen3.6-35B-A3B-NVFP4 NVFP4", "short_name": "Qwen3.6 35B A3B NVFP4", "license": "apache-2.0", "hf_pipeline_tag": "unknown", "hf_downloads": 2262556, "tags": [ "safetensors", "qwen3_5_moe", "qwen", "nvfp4", "vllm", "compressed-tensors", "base_model:Qwen/Qwen3.6-35B-A3B", "base_model:quantized:Qwen/Qwen3.6-35B-A3B", "license:apache-2.0", "8-bit" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_5_moe_text", "total_params_b": 35.9518, "active_params_b": 3, "layers": 40, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "NVFP4", "weight_bytes_per_param": 0.5, "resident_weight_gb": 17.9759, "active_weight_gb": 1.5, "kv_alloc_gb_per_1k_tokens": 0.08192, "kv_read_gb_per_1k_tokens": 0.08192, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=2262556; data quality: config, params, base:Qwen/Qwen3.6-35B-A3B. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/RedHatAI/Qwen3.6-35B-A3B-NVFP4" }, { "label": "Model config", "url": "https://huggingface.co/RedHatAI/Qwen3.6-35B-A3B-NVFP4/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B" } ] }, { "id": "Qwen/Qwen2.5-14B-Instruct", "name": "Qwen2.5-14B-Instruct BF16", "short_name": "Qwen2.5 14B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 2124618, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "chat", "conversational", "en", "arxiv:2309.00071", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-14B", "base_model:finetune:Qwen/Qwen2.5-14B", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 14.77, "active_params_b": 14.77, "layers": 48, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 29.54, "active_weight_gb": 29.54, "kv_alloc_gb_per_1k_tokens": 0.196608, "kv_read_gb_per_1k_tokens": 0.196608, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=2124618; data quality: config, params, base:Qwen/Qwen2.5-14B. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-14B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-14B-Instruct/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-14B" } ] }, { "id": "Qwen/Qwen3-VL-2B-Instruct", "name": "Qwen3-VL-2B-Instruct BF16", "short_name": "Qwen3 VL 2B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 2116507, "tags": [ "transformers", "safetensors", "qwen3_vl", "image-text-to-text", "conversational", "arxiv:2505.09388", "arxiv:2502.13923", "arxiv:2409.12191", "arxiv:2308.12966", "license:apache-2.0", "eval-results", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_vl_text", "total_params_b": 2.1275, "active_params_b": 2.1275, "layers": 28, "attention_heads": 16, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 4.255, "active_weight_gb": 4.255, "kv_alloc_gb_per_1k_tokens": 0.114688, "kv_read_gb_per_1k_tokens": 0.114688, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=2116507; data quality: config, params. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-2B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-VL-2B-Instruct/raw/main/config.json" } ] }, { "id": "nvidia/Gemma-4-26B-A4B-NVFP4", "name": "Gemma-4-26B-A4B-NVFP4 NVFP4", "short_name": "Gemma 4 26B A4B NVFP4", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 2107107, "tags": [ "Model Optimizer", "safetensors", "gemma4", "nvidia", "ModelOpt", "quantized", "NVFP4", "nvfp4", "gemma4-26b-A4B-it", "text-generation", "conversational", "base_model:google/gemma-4-26B-A4B-it", "base_model:quantized:google/gemma-4-26B-A4B-it", "license:apache-2.0", "8-bit", "modelopt" ], "architecture": { "type": "MoE transformer", "detail": "gemma4_text", "total_params_b": 26.5441, "active_params_b": 4, "layers": 30, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 1, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "NVFP4", "weight_bytes_per_param": 0.5, "resident_weight_gb": 13.2721, "active_weight_gb": 2, "kv_alloc_gb_per_1k_tokens": 0.10496, "kv_read_gb_per_1k_tokens": 0.10496, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=2107107; data quality: config, params, base:google/gemma-4-26B-A4B-it. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/Gemma-4-26B-A4B-NVFP4" }, { "label": "Model config", "url": "https://huggingface.co/nvidia/Gemma-4-26B-A4B-NVFP4/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it" } ] }, { "id": "meta-llama/Llama-3.2-3B-Instruct", "name": "Llama-3.2-3B-Instruct BF16", "short_name": "Llama 3.2 3B", "license": "llama3.2", "hf_pipeline_tag": "text-generation", "hf_downloads": 2056857, "tags": [ "transformers", "safetensors", "llama", "text-generation", "facebook", "meta", "pytorch", "llama-3", "conversational", "en", "de", "fr", "it", "pt", "hi", "es", "th", "arxiv:2204.05149", "arxiv:2405.16406", "license:llama3.2", "text-generation-inference", "endpoints_compatible" ], "architecture": { "type": "Dense transformer", "detail": "unknown", "total_params_b": 3.2127, "active_params_b": 3.2127, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": 2048, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 6.4254, "active_weight_gb": 6.4254, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=2056857; data quality: no-config, params. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct" } ] }, { "id": "Qwen/Qwen3-TTS-12Hz-1.7B-Base", "name": "Qwen3-TTS-12Hz-1.7B-Base BF16", "short_name": "Qwen3 TTS 12Hz 1.7B", "license": "apache-2.0", "hf_pipeline_tag": "unknown", "hf_downloads": 2047748, "tags": [ "safetensors", "qwen3_tts", "arxiv:2601.15621", "license:apache-2.0" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_tts (Qwen3TTSForConditionalGeneration)", "total_params_b": 1.9287, "active_params_b": 1.9287, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": 2048, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 3.8574, "active_weight_gb": 3.8574, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=2047748; data quality: config, params. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-Base" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-Base/raw/main/config.json" } ] }, { "id": "Qwen/Qwen3-0.6B-FP8", "name": "Qwen3-0.6B-FP8 FP8", "short_name": "Qwen3 0.6B FP8", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 1984605, "tags": [ "transformers", "safetensors", "qwen3", "text-generation", "conversational", "arxiv:2505.09388", "base_model:Qwen/Qwen3-0.6B", "base_model:quantized:Qwen/Qwen3-0.6B", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "fp8" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 0.7517, "active_params_b": 0.7517, "layers": 28, "attention_heads": 16, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "FP8", "weight_bytes_per_param": 1, "resident_weight_gb": 0.7517, "active_weight_gb": 0.7517, "kv_alloc_gb_per_1k_tokens": 0.114688, "kv_read_gb_per_1k_tokens": 0.114688, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=1984605; data quality: config, params, base:Qwen/Qwen3-0.6B. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-0.6B-FP8" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-0.6B-FP8/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-0.6B" } ] }, { "id": "google/gemma-4-12B-it-qat-w4a16-ct", "name": "gemma-4-12B-it-qat-w4a16-ct 4-bit", "short_name": "gemma 4 12B it qat w4a16 ct", "license": "apache-2.0", "hf_pipeline_tag": "any-to-any", "hf_downloads": 1963007, "tags": [ "transformers", "safetensors", "gemma4_unified", "image-text-to-text", "any-to-any", "base_model:google/gemma-4-12B-it-qat-q4_0-unquantized", "base_model:quantized:google/gemma-4-12B-it-qat-q4_0-unquantized", "license:apache-2.0", "endpoints_compatible", "compressed-tensors" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_unified_text", "total_params_b": 13.307, "active_params_b": 13.307, "layers": 48, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "4-bit", "weight_bytes_per_param": 0.5, "resident_weight_gb": 6.6535, "active_weight_gb": 6.6535, "kv_alloc_gb_per_1k_tokens": 0.393216, "kv_read_gb_per_1k_tokens": 0.393216, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=1963007; data quality: config, params, base:google/gemma-4-12B-it-qat-q4_0-unquantized. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-4-12B-it-qat-w4a16-ct" }, { "label": "Model config", "url": "https://huggingface.co/google/gemma-4-12B-it-qat-w4a16-ct/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-12B-it-qat-q4_0-unquantized" } ] }, { "id": "nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4", "name": "Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4 NVFP4", "short_name": "Nemotron 3 Nano Omni 30B A3B Reasoning NVFP4", "license": "other", "hf_pipeline_tag": "any-to-any", "hf_downloads": 1961825, "tags": [ "transformers", "safetensors", "NemotronH_Nano_Omni_Reasoning_V3", "feature-extraction", "nvidia", "pytorch", "multimodal", "any-to-any", "custom_code", "dataset:nvidia/Nemotron-Image-Training-v3", "arxiv:2604.24954", "base_model:nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16", "base_model:quantized:nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16", "license:other", "8-bit", "modelopt", "deploy:azure" ], "architecture": { "type": "MoE transformer", "detail": "nemotron_h (NemotronHForCausalLM)", "total_params_b": 33.0156, "active_params_b": 3, "layers": 52, "attention_heads": 32, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 6, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "NVFP4", "weight_bytes_per_param": 0.5, "resident_weight_gb": 16.5078, "active_weight_gb": 1.5, "kv_alloc_gb_per_1k_tokens": 0.053248, "kv_read_gb_per_1k_tokens": 0.053248, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=1961825; data quality: config, params, base:nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4" }, { "label": "Model config", "url": "https://huggingface.co/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16" } ] }, { "id": "Qwen/Qwen3.5-35B-A3B", "name": "Qwen3.5-35B-A3B BF16/F32", "short_name": "Qwen3.5 35B A3B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1955420, "tags": [ "transformers", "safetensors", "qwen3_5_moe", "image-text-to-text", "conversational", "base_model:Qwen/Qwen3.5-35B-A3B-Base", "base_model:finetune:Qwen/Qwen3.5-35B-A3B-Base", "license:apache-2.0", "eval-results", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_5_moe_text", "total_params_b": 35.9518, "active_params_b": 3, "layers": 40, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "BF16/F32", "weight_bytes_per_param": 2, "resident_weight_gb": 71.9036, "active_weight_gb": 6, "kv_alloc_gb_per_1k_tokens": 0.08192, "kv_read_gb_per_1k_tokens": 0.08192, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=1955420; data quality: config, params, base:Qwen/Qwen3.5-35B-A3B-Base. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B-Base" } ] }, { "id": "Qwen/Qwen3-Coder-Next-FP8", "name": "Qwen3-Coder-Next-FP8 FP8", "short_name": "Qwen3 Coder Next FP8", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 1955309, "tags": [ "transformers", "safetensors", "qwen3_next", "text-generation", "conversational", "license:apache-2.0", "endpoints_compatible", "fp8", "deploy:azure" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_next (Qwen3NextForCausalLM)", "total_params_b": 79.6792, "active_params_b": 1.5562, "layers": 48, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "FP8", "weight_bytes_per_param": 1, "resident_weight_gb": 79.6792, "active_weight_gb": 1.5562, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=1955309; data quality: config, params. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-Coder-Next-FP8" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-Coder-Next-FP8/raw/main/config.json" } ] }, { "id": "Qwen/Qwen2.5-Coder-7B-Instruct", "name": "Qwen2.5-Coder-7B-Instruct BF16", "short_name": "Qwen2.5 Coder 7B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 1935745, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "code", "codeqwen", "chat", "qwen", "qwen-coder", "conversational", "en", "arxiv:2409.12186", "arxiv:2309.00071", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-Coder-7B", "base_model:finetune:Qwen/Qwen2.5-Coder-7B", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 7.6156, "active_params_b": 7.6156, "layers": 28, "attention_heads": 28, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 15.2312, "active_weight_gb": 15.2312, "kv_alloc_gb_per_1k_tokens": 0.057344, "kv_read_gb_per_1k_tokens": 0.057344, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=1935745; data quality: config, params, base:Qwen/Qwen2.5-Coder-7B. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B" } ] }, { "id": "nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16", "name": "NVIDIA-Nemotron-3-Nano-4B-BF16 BF16", "short_name": "NVIDIA Nemotron 3 Nano 4B BF16", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 1929605, "tags": [ "transformers", "safetensors", "nemotron_h", "text-generation", "nvidia", "pytorch", "conversational", "custom_code", "en", "dataset:nvidia/Nemotron-CC-v2", "dataset:nvidia/Nemotron-Post-Training-Dataset-v2", "dataset:nvidia/Nemotron-Science-v1", "dataset:nvidia/Nemotron-Instruction-Following-Chat-v1", "dataset:nvidia/Nemotron-Agentic-v1", "dataset:nvidia/Nemotron-Competitive-Programming-v1", "dataset:nvidia/Nemotron-Math-Proofs-v1", "dataset:nvidia/Nemotron-RL-Agentic-Conversational-Tool-Use-Pivot-v1", "dataset:nvidia/Nemotron-RL-instruction_following", "dataset:nvidia/Nemotron-RL-agent-calendar_scheduling", "dataset:nvidia/Nemotron-RL-instruction_following-structured_outputs", "arxiv:2511.16664", "arxiv:2504.03624", "arxiv:2512.20856", "arxiv:2512.20848", "arxiv:2412.02595", "base_model:nvidia/NVIDIA-Nemotron-Nano-9B-v2", "base_model:finetune:nvidia/NVIDIA-Nemotron-Nano-9B-v2", "license:other", "endpoints_compatible" ], "architecture": { "type": "Hybrid Mamba/Transformer", "detail": "nemotron_h (NemotronHForCausalLM); 21 Mamba, 4 attention, 17 MLP layers", "total_params_b": 3.9736, "active_params_b": 3.5625, "layers": 42, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "hybrid", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 7.9471, "active_weight_gb": 7.125, "kv_alloc_gb_per_1k_tokens": 0.016384, "kv_read_gb_per_1k_tokens": 0.016384, "kv_alloc_fixed_gb": 0.042922, "kv_read_fixed_gb": 0.042922, "notes": "Manually corrected on 2026-07-02 from current Hugging Face API metadata, model card, served config, custom runtime code, and safetensors headers. The repo currently records BF16:3973556832, a 42-layer hybrid Mamba2/Transformer pattern with 21 Mamba layers, 4 full-attention layers, and 17 MLP layers. The audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16" }, { "label": "Model config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2" } ] }, { "id": "Qwen/Qwen3-Omni-30B-A3B-Instruct", "name": "Qwen3-Omni-30B-A3B-Instruct BF16", "short_name": "Qwen3 Omni 30B A3B", "license": "other", "hf_pipeline_tag": "any-to-any", "hf_downloads": 1894886, "tags": [ "transformers", "safetensors", "qwen3_omni_moe", "text-to-audio", "multimodal", "any-to-any", "en", "license:other", "endpoints_compatible" ], "architecture": { "type": "Multimodal MoE transformer", "detail": "qwen3_omni_moe Thinker/Talker package; ordinary text decode uses the 48-layer Thinker text MoE", "total_params_b": 35.2598, "active_params_b": 3.0421, "layers": 48, "attention_heads": 32, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 65536, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 70.51963709, "active_weight_gb": 6.08425984, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "expert_param_b": 0.226492416, "fixed_param_b": 1.230190592, "notes": "Manual correction on 2026-07-02 from served config and range-read safetensors headers. Full Instruct package is resident; ordinary text decode sweeps only Thinker text non-embedding weights and expected distinct routed experts. Profile data supersedes this legacy adapter for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Instruct/raw/main/config.json" } ] }, { "id": "deepseek-ai/DeepSeek-V4-Flash", "name": "DeepSeek-V4-Flash FP4/FP8", "short_name": "DeepSeek V4 Flash", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 1935113, "tags": [ "transformers", "safetensors", "deepseek_v4", "text-generation", "conversational", "arxiv:2606.19348", "license:mit", "eval-results", "endpoints_compatible", "8-bit", "fp8", "deploy:azure", "region:us" ], "architecture": { "type": "Mixed FP4/FP8 MoE transformer", "detail": "deepseek_v4 CSA/HCA MoE (DeepseekV4ForCausalLM)", "total_params_b": 284, "active_params_b": 13, "layers": 43, "attention_heads": 64, "kv_heads": 1, "head_dim": 512, "max_context_tokens": 1048576, "routed_experts": 256, "routed_experts_per_token": 6, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "FP4/FP8 mixed", "weight_bytes_per_param": 1, "resident_weight_gb": 159.609485896, "active_weight_gb": 11.23618838, "kv_alloc_gb_per_1k_tokens": 0.00688, "kv_read_gb_per_1k_tokens": 0.001504, "kv_alloc_fixed_gb": 0.017842176, "kv_read_fixed_gb": 0.016646144, "expert_param_b": 0.574881792, "shared_expert_param_b": 1.08219648, "fixed_param_b": 6.704701148, "notes": "Manual correction on 2026-07-02 from model card, served config, official inference code, and range-read safetensors headers at SHA 60d8d70770c6776ff598c94bb586a859a38244f1. Catalog parameters use the model-card logical 284B total / 13B active values; bound weights use exact stored FP4/FP8/BF16/F32/I64 bytes. Profile data supersedes this legacy adapter for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash" }, { "label": "Model config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash/raw/main/config.json" } ] }, { "id": "meta-llama/Llama-3.2-1B", "name": "Llama-3.2-1B BF16", "short_name": "Llama 3.2 1B", "license": "llama3.2", "hf_pipeline_tag": "text-generation", "hf_downloads": 1878682, "tags": [ "transformers", "safetensors", "llama", "text-generation", "facebook", "meta", "pytorch", "llama-3", "en", "de", "fr", "it", "pt", "hi", "es", "th", "arxiv:2204.05149", "arxiv:2405.16406", "license:llama3.2", "eval-results", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "llama3.2 dense decoder (gated config not directly audited)", "total_params_b": 1.2358144, "active_params_b": 1.2358144, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": 128000, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 2.4716288, "active_weight_gb": 2.4716288, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manual correction on 2026-07-02 from Hugging Face API/model card metadata. API reports gated: manual and BF16 safetensors count 1235814400; public card reports 128k context, GQA, and shared embeddings. Raw config and weights are not accessible in this audit environment, so the unsupported profile disables production bounds instead of using this legacy zero-KV estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/meta-llama/Llama-3.2-1B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.2-1B" } ] }, { "id": "RedHatAI/gemma-4-26B-A4B-it-FP8-Dynamic", "name": "gemma-4-26B-A4B-it-FP8-Dynamic FP8", "short_name": "gemma 4 26B A4B it FP8 Dynamic", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1931625, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "fp8", "vllm", "llm-compressor", "compressed-tensors", "conversational", "base_model:google/gemma-4-26B-A4B-it", "base_model:quantized:google/gemma-4-26B-A4B-it", "license:apache-2.0", "endpoints_compatible" ], "architecture": { "type": "MoE transformer", "detail": "gemma4_text", "total_params_b": 26.5609, "active_params_b": 4, "layers": 30, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "FP8", "weight_bytes_per_param": 1, "resident_weight_gb": 28.63837046, "active_weight_gb": 4.57536262, "kv_alloc_gb_per_1k_tokens": 0.20992, "kv_read_gb_per_1k_tokens": 0.20992, "notes": "Manually corrected on 2026-07-02 from current Hugging Face API metadata, model card, served config, Google base config comparison, and direct safetensors header range reads. KV display fields show first-1K text-cache traffic; the audited profile supersedes these legacy adapter fields for production bounds and handles the 1024-token sliding-window cap." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/RedHatAI/gemma-4-26B-A4B-it-FP8-Dynamic" }, { "label": "Model config", "url": "https://huggingface.co/RedHatAI/gemma-4-26B-A4B-it-FP8-Dynamic/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/RedHatAI/gemma-4-26B-A4B-it-FP8-Dynamic" } ] }, { "id": "cyankiwi/Qwen3.6-35B-A3B-AWQ-4bit", "name": "Qwen3.6-35B-A3B-AWQ-4bit 4-bit", "short_name": "Qwen3.6 35B A3B AWQ 4bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1845264, "tags": [ "transformers", "safetensors", "qwen3_5_moe", "image-text-to-text", "conversational", "base_model:Qwen/Qwen3.6-35B-A3B", "base_model:quantized:Qwen/Qwen3.6-35B-A3B", "license:apache-2.0", "endpoints_compatible", "4-bit", "awq", "deploy:azure" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_5_moe_text", "total_params_b": 35.9518, "active_params_b": 3, "layers": 40, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "4-bit", "weight_bytes_per_param": 0.5, "resident_weight_gb": 24.956800736, "active_weight_gb": 4.461552896, "kv_alloc_gb_per_1k_tokens": 0.08536064, "kv_read_gb_per_1k_tokens": 0.08536064, "notes": "Manually corrected on 2026-07-02 from current Hugging Face API metadata, model card, served config, base config comparison, quantization config, and direct safetensors header range reads. KV display fields show first-1K text-cache plus fixed DeltaNet state traffic; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/cyankiwi/Qwen3.6-35B-A3B-AWQ-4bit" }, { "label": "Model config", "url": "https://huggingface.co/cyankiwi/Qwen3.6-35B-A3B-AWQ-4bit/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3.6-35B-A3B-AWQ-4bit" } ] }, { "id": "Qwen/Qwen2-VL-7B-Instruct", "name": "Qwen2-VL-7B-Instruct BF16", "short_name": "Qwen2 VL 7B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1797290, "tags": [ "transformers", "safetensors", "qwen2_vl", "image-text-to-text", "multimodal", "conversational", "en", "arxiv:2409.12191", "arxiv:2308.12966", "base_model:Qwen/Qwen2-VL-7B", "base_model:finetune:Qwen/Qwen2-VL-7B", "license:apache-2.0", "eval-results", "text-generation-inference", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen2_vl (Qwen2VLForConditionalGeneration)", "total_params_b": 8.291375616, "active_params_b": 7.070619136, "layers": 28, "attention_heads": 28, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 16.582751232, "active_weight_gb": 14.141238272, "kv_alloc_gb_per_1k_tokens": 0.057344, "kv_read_gb_per_1k_tokens": 0.057344, "notes": "Manually corrected on 2026-07-02 from current Hugging Face API metadata, model card, served config, base config comparison, and direct safetensors header range reads. Bound weights separate the 16.582751232 GB resident multimodal package from the 14.141238272 GB ordinary text-decode swept subset; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2-VL-7B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2-VL-7B-Instruct" } ] }, { "id": "Qwen/Qwen2-VL-7B-Instruct-AWQ", "name": "Qwen2-VL-7B-Instruct-AWQ 4-bit", "short_name": "Qwen2 VL 7B Instruct", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1806813, "tags": [ "transformers", "safetensors", "qwen2_vl", "image-text-to-text", "multimodal", "conversational", "en", "arxiv:2409.12191", "arxiv:2308.12966", "base_model:Qwen/Qwen2-VL-7B-Instruct", "base_model:quantized:Qwen/Qwen2-VL-7B-Instruct", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "4-bit", "awq", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen2_vl (Qwen2VLForConditionalGeneration)", "total_params_b": 8.291375616, "active_params_b": 7.070619136, "layers": 28, "attention_heads": 28, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "4-bit", "weight_bytes_per_param": 0.5, "resident_weight_gb": 6.9222656, "active_weight_gb": 4.48075264, "kv_alloc_gb_per_1k_tokens": 0.057344, "kv_read_gb_per_1k_tokens": 0.057344, "notes": "Manually corrected on 2026-07-02 from current Hugging Face API metadata, model card, served AWQ config, BF16 base config comparison, and direct safetensors header range reads. Bound weights separate the 6.9222656 GB resident multimodal AWQ package from the 4.48075264 GB ordinary text-decode swept subset; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct-AWQ" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct-AWQ/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2-VL-7B-Instruct-AWQ" } ] }, { "id": "RedHatAI/Llama-3.2-1B-Instruct-FP8-dynamic", "name": "Llama-3.2-1B-Instruct-FP8-dynamic FP8", "short_name": "Llama 3.2 1B Instruct FP8 dynamic", "license": "llama3.2", "hf_pipeline_tag": "text-generation", "hf_downloads": 1792332, "tags": [ "safetensors", "llama", "fp8", "vllm", "text-generation", "conversational", "en", "de", "fr", "it", "pt", "hi", "es", "th", "base_model:meta-llama/Llama-3.2-1B-Instruct", "base_model:quantized:meta-llama/Llama-3.2-1B-Instruct", "license:llama3.2", "compressed-tensors" ], "architecture": { "type": "Dense transformer", "detail": "llama (LlamaForCausalLM)", "total_params_b": 1.49885952, "active_params_b": 1.236191232, "layers": 16, "attention_heads": 32, "kv_heads": 8, "head_dim": 64, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "FP8", "weight_bytes_per_param": 1, "resident_weight_gb": 2.024640512, "active_weight_gb": 1.499303936, "kv_alloc_gb_per_1k_tokens": 0.032768, "kv_read_gb_per_1k_tokens": 0.032768, "notes": "Manually corrected on 2026-07-02 from current Hugging Face API metadata, model card, served FP8 dynamic config, and direct safetensors header range reads. Bound weights separate the 2.024640512 GB resident compressed-tensors package from the 1.499303936 GB ordinary text-decode swept subset; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/RedHatAI/Llama-3.2-1B-Instruct-FP8-dynamic" }, { "label": "Model config", "url": "https://huggingface.co/RedHatAI/Llama-3.2-1B-Instruct-FP8-dynamic/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/RedHatAI/Llama-3.2-1B-Instruct-FP8-dynamic" } ] }, { "id": "google/gemma-3-12b-it", "name": "gemma-3-12b-it BF16", "short_name": "gemma 3 12b it", "license": "gemma", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1705245, "tags": [ "transformers", "safetensors", "gemma3", "image-text-to-text", "conversational", "arxiv:1905.07830", "arxiv:1905.10044", "arxiv:1911.11641", "arxiv:1904.09728", "arxiv:1705.03551", "arxiv:1911.01547", "arxiv:1907.10641", "arxiv:1903.00161", "arxiv:2009.03300", "arxiv:2304.06364", "arxiv:2103.03874", "arxiv:2110.14168", "arxiv:2311.12022", "arxiv:2108.07732", "arxiv:2107.03374", "arxiv:2210.03057", "arxiv:2106.03193", "arxiv:1910.11856", "arxiv:2502.12404", "arxiv:2502.21228", "arxiv:2404.16816", "arxiv:2104.12756", "arxiv:2311.16502", "arxiv:2203.10244", "arxiv:2404.12390", "arxiv:1810.12440", "arxiv:1908.02660", "arxiv:2312.11805", "base_model:google/gemma-3-12b-pt", "base_model:finetune:google/gemma-3-12b-pt", "license:gemma", "eval-results", "text-generation-inference", "endpoints_compatible" ], "architecture": { "type": "Dense transformer", "detail": "unknown", "total_params_b": 12.18732504, "active_params_b": 12.18732504, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": 2048, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 24.37465008, "active_weight_gb": 24.37465008, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manually corrected on 2026-07-02 from current public Hugging Face API metadata. Production bounds are disabled by an unsupported profile because raw config, README, generation config, safetensors index, safetensors bytes, and hf download config.json are gated in this audit environment." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-3-12b-it" }, { "label": "Base model config", "url": "https://huggingface.co/google/gemma-3-12b-pt/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-3-12b-pt" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-3-12b-it" } ] }, { "id": "Qwen/Qwen3-Coder-30B-A3B-Instruct", "name": "Qwen3-Coder-30B-A3B-Instruct BF16", "short_name": "Qwen3 Coder 30B A3B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 1714841, "tags": [ "transformers", "safetensors", "qwen3_moe", "text-generation", "conversational", "arxiv:2505.09388", "license:apache-2.0", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_moe (Qwen3MoeForCausalLM)", "total_params_b": 30.532122624, "active_params_b": 3.041867776, "layers": 48, "attention_heads": 32, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 61.064245248, "active_weight_gb": 6.083735552, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "notes": "Manually corrected on 2026-07-02 from current Hugging Face API metadata, model card, served config, safetensors index, and direct safetensors header range reads. Active traffic is fixed decode traffic plus 8 routed expert groups; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct/raw/main/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-Coder-30B-A3B-Instruct" } ] }, { "id": "MiniMaxAI/MiniMax-M2.7", "name": "MiniMax-M2.7 FP8", "short_name": "MiniMax M2.7", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 1596931, "tags": [ "transformers", "safetensors", "minimax_m2", "text-generation", "conversational", "custom_code", "license:other", "eval-results", "endpoints_compatible", "fp8" ], "architecture": { "type": "MoE transformer", "detail": "minimax_m2 (MiniMaxM2ForCausalLM)", "total_params_b": 228.703644928, "active_params_b": 10.41655168, "layers": 62, "attention_heads": 48, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 204800, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "FP8", "weight_bytes_per_param": 1, "resident_weight_gb": 230.12163072, "active_weight_gb": 11.180086272, "kv_alloc_gb_per_1k_tokens": 0.253952, "kv_read_gb_per_1k_tokens": 0.253952, "notes": "Manually corrected on 2026-07-02 from current Hugging Face API metadata, model card, served config, custom model code, safetensors index, and direct safetensors header range reads. Active traffic is fixed decode traffic plus 8 routed expert groups; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M2.7" }, { "label": "Model config", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M2.7/raw/main/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MiniMaxAI/MiniMax-M2.7" } ] }, { "id": "google/gemma-3-4b-it", "name": "gemma-3-4b-it BF16", "short_name": "gemma 3 4b it", "license": "gemma", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1795040, "tags": [ "transformers", "safetensors", "gemma3", "image-text-to-text", "conversational", "arxiv:1905.07830", "arxiv:1905.10044", "arxiv:1911.11641", "arxiv:1904.09728", "arxiv:1705.03551", "arxiv:1911.01547", "arxiv:1907.10641", "arxiv:1903.00161", "arxiv:2009.03300", "arxiv:2304.06364", "arxiv:2103.03874", "arxiv:2110.14168", "arxiv:2311.12022", "arxiv:2108.07732", "arxiv:2107.03374", "arxiv:2210.03057", "arxiv:2106.03193", "arxiv:1910.11856", "arxiv:2502.12404", "arxiv:2502.21228", "arxiv:2404.16816", "arxiv:2104.12756", "arxiv:2311.16502", "arxiv:2203.10244", "arxiv:2404.12390", "arxiv:1810.12440", "arxiv:1908.02660", "arxiv:2312.11805", "base_model:google/gemma-3-4b-pt", "base_model:finetune:google/gemma-3-4b-pt", "license:gemma", "eval-results", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense multimodal transformer", "detail": "Gemma3ForConditionalGeneration (gated raw config)", "total_params_b": 4.300079472, "active_params_b": 4.300079472, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 8.600158944, "active_weight_gb": 8.600158944, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manually corrected on 2026-07-02 from current Hugging Face API metadata. The repo is gated; raw config, README, generation config, safetensors index, ranged safetensors bytes, and hf download config.json all returned access denied. Production bounds use the unsupported profile and must not infer KV geometry from this row." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-3-4b-it" }, { "label": "Base model config", "url": "https://huggingface.co/google/gemma-3-4b-pt/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-3-4b-pt" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-3-4b-it" } ] }, { "id": "Qwen/Qwen2.5-Omni-3B", "name": "Qwen2.5-Omni-3B F32/BF16", "short_name": "Qwen2.5 Omni 3B", "license": "other", "hf_pipeline_tag": "any-to-any", "hf_downloads": 1736821, "tags": [ "transformers", "safetensors", "qwen2_5_omni", "multimodal", "any-to-any", "en", "arxiv:2503.20215", "license:other", "endpoints_compatible", "region:us" ], "architecture": { "type": "Multimodal dense transformer", "detail": "qwen2_5_omni Thinker/Talker package; ordinary text decode uses the 36-layer Thinker text decoder", "total_params_b": 5.537120672, "active_params_b": 3.085938688, "layers": 36, "attention_heads": 16, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "F32/BF16", "weight_bytes_per_param": 2.162196680404945, "resident_weight_gb": 11.972343936, "active_weight_gb": 6.171877376, "kv_alloc_gb_per_1k_tokens": 0.036864, "kv_read_gb_per_1k_tokens": 0.036864, "notes": "Manual correction on 2026-07-02 from served config, model card, Hugging Face API metadata, safetensors index, and range-read shard headers. Full Thinker/Talker/Token2Wav checkpoint is resident; ordinary text decode sweeps only the Thinker text decoder plus lm_head. Profile data supersedes this legacy adapter for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Omni-3B" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-Omni-3B/raw/main/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Omni-3B" } ] }, { "id": "Qwen/Qwen3-14B-AWQ", "name": "Qwen3-14B-AWQ 4-bit", "short_name": "Qwen3 14B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 1699790, "tags": [ "transformers", "safetensors", "qwen3", "text-generation", "conversational", "arxiv:2309.00071", "arxiv:2505.09388", "base_model:Qwen/Qwen3-14B", "base_model:quantized:Qwen/Qwen3-14B", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "4-bit", "awq", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 14.7683072, "active_params_b": 13.99039488, "layers": 40, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "4-bit", "weight_bytes_per_param": 0.5, "resident_weight_gb": 9.976576, "active_weight_gb": 8.42075136, "kv_alloc_gb_per_1k_tokens": 0.16384, "kv_read_gb_per_1k_tokens": 0.16384, "notes": "Manual correction on 2026-07-02 from current Hugging Face API metadata, model card, served AWQ config, base Qwen3-14B profile comparison, safetensors index, and direct safetensors header range reads. Bound weights separate the 9.976576 GB resident AWQ package from the 8.42075136 GB ordinary text-decode swept subset; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-14B-AWQ" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-14B-AWQ/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-14B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-14B-AWQ" } ] }, { "id": "Qwen/Qwen2.5-Coder-32B-Instruct-AWQ", "name": "Qwen2.5-Coder-32B-Instruct-AWQ 4-bit", "short_name": "Qwen2.5 Coder 32B Instruct", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 1707036, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "code", "codeqwen", "chat", "qwen", "qwen-coder", "conversational", "en", "arxiv:2409.12186", "arxiv:2309.00071", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-Coder-32B-Instruct", "base_model:quantized:Qwen/Qwen2.5-Coder-32B-Instruct", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "4-bit", "awq", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 32.763876352, "active_params_b": 31.985308672, "layers": 64, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "4-bit", "weight_bytes_per_param": 0.5, "resident_weight_gb": 19.328804864, "active_weight_gb": 17.771669504, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Manual correction on 2026-07-02 from current Hugging Face API metadata, model card, served AWQ config, base Qwen2.5-Coder 32B Instruct profile comparison, safetensors index, and direct safetensors header range reads. Bound weights separate the 19.328804864 GB resident AWQ package from the 17.771669504 GB ordinary text-decode swept subset; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct-AWQ" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct-AWQ/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Coder-32B-Instruct-AWQ" } ] }, { "id": "Qwen/Qwen3-VL-235B-A22B-Instruct", "name": "Qwen3-VL-235B-A22B-Instruct BF16", "short_name": "Qwen3 VL 235B A22B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1676772, "tags": [ "transformers", "safetensors", "qwen3_vl_moe", "image-text-to-text", "conversational", "arxiv:2505.09388", "arxiv:2502.13923", "arxiv:2409.12191", "arxiv:2308.12966", "license:apache-2.0", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_vl_moe_text", "total_params_b": 235.670022896, "active_params_b": 21.568433664, "layers": 94, "attention_heads": 64, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 471.340045792, "active_weight_gb": 43.136867328, "kv_alloc_gb_per_1k_tokens": 0.192512, "kv_read_gb_per_1k_tokens": 0.192512, "notes": "Manual correction on 2026-07-02 from current Hugging Face API metadata, served config, model card, safetensors index, and direct safetensors header range reads. Bound weights separate the 471.340045792 GB resident BF16 package from the 43.136867328 GB ordinary text-decode active traffic; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Instruct/raw/main/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-VL-235B-A22B-Instruct" }, { "label": "Safetensors index", "url": "https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Instruct/raw/710c13861be6c466e66de3f484069440b8f31389/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen3.5-2B", "name": "Qwen3.5-2B F32/BF16", "short_name": "Qwen3.5 2B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1706576, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "conversational", "base_model:Qwen/Qwen3.5-2B-Base", "base_model:finetune:Qwen/Qwen3.5-2B-Base", "license:apache-2.0", "eval-results", "endpoints_compatible", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 2.274069824, "active_params_b": 1.881825088, "layers": 24, "attention_heads": 8, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "F32/BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 4.548144832, "active_weight_gb": 3.76365536, "kv_alloc_gb_per_1k_tokens": 0.049152, "kv_read_gb_per_1k_tokens": 0.049152, "notes": "Manual correction on 2026-07-02 from current Hugging Face API metadata, served config, base config comparison, model card, safetensors index, and direct safetensors header range reads. Bound weights separate the 4.548144832 GB resident BF16/F32 package from the 3.76365536 GB ordinary text-decode swept subset; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3.5-2B" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3.5-2B/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-2B-Base" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-2B" }, { "label": "Safetensors index", "url": "https://huggingface.co/Qwen/Qwen3.5-2B/raw/15852e8c16360a2fea060d615a32b45270f8a8fc/model.safetensors.index.json" } ] }, { "id": "moonshotai/Kimi-K2-Instruct-0905", "name": "Kimi-K2-Instruct-0905 FP8", "short_name": "Kimi K2 Instruct 0905", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 1591378, "tags": [ "transformers", "safetensors", "kimi_k2", "text-generation", "conversational", "custom_code", "license:other", "eval-results", "endpoints_compatible", "fp8" ], "architecture": { "type": "MoE transformer", "detail": "kimi_k2 (DeepseekV3ForCausalLM)", "total_params_b": 1026.470735448, "active_params_b": 31.688957528, "layers": 61, "attention_heads": 64, "kv_heads": 64, "head_dim": 112, "max_context_tokens": 262144, "routed_experts": 384, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "FP8", "weight_bytes_per_param": 1, "resident_weight_gb": 1029.173265504, "active_weight_gb": 33.035158624, "kv_alloc_gb_per_1k_tokens": 1.748992, "kv_read_gb_per_1k_tokens": 1.748992, "notes": "Manual correction on 2026-07-02 from current Hugging Face API metadata, served FP8 config, model card, deployment guide, custom modeling code, safetensors index, and direct safetensors header range reads. Bound weights separate the 1029.173265504 GB resident block-FP8 package from the 33.035158624 GB ordinary text-decode active traffic; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/moonshotai/Kimi-K2-Instruct-0905" }, { "label": "Model config", "url": "https://huggingface.co/moonshotai/Kimi-K2-Instruct-0905/raw/main/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/moonshotai/Kimi-K2-Instruct-0905" }, { "label": "Safetensors index", "url": "https://huggingface.co/moonshotai/Kimi-K2-Instruct-0905/raw/ac6c49f04883bd0a0598b790693a72061c676629/model.safetensors.index.json" } ] }, { "id": "lmstudio-community/gemma-4-E4B-it-MLX-4bit", "name": "gemma-4-E4B-it-MLX-4bit 4-bit", "short_name": "gemma 4 E4B it MLX 4bit", "license": "apache-2.0", "hf_pipeline_tag": "any-to-any", "hf_downloads": 1681799, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "mlx", "any-to-any", "base_model:google/gemma-4-E4B-it", "base_model:quantized:google/gemma-4-E4B-it", "license:apache-2.0", "endpoints_compatible", "4-bit" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 7.996157418, "active_params_b": 7.996157418, "layers": 42, "attention_heads": 8, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "MLX 4-bit affine", "weight_bytes_per_param": 0.8540267149603031, "resident_weight_gb": 6.828932052, "active_weight_gb": 4.295787092, "kv_alloc_gb_per_1k_tokens": 0.086016, "kv_read_gb_per_1k_tokens": 0.086016, "notes": "Manual correction on 2026-07-02 from current Hugging Face API metadata, served MLX config, base Google config comparison, model card, safetensors index, and direct shard header range reads. Bound weights separate the exact 6.828932052 GB resident MLX package from the 4.295787092 GB ordinary text-decode swept traffic; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/gemma-4-E4B-it-MLX-4bit" }, { "label": "Model config", "url": "https://huggingface.co/lmstudio-community/gemma-4-E4B-it-MLX-4bit/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-E4B-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-E4B-it-MLX-4bit" }, { "label": "Safetensors index", "url": "https://huggingface.co/lmstudio-community/gemma-4-E4B-it-MLX-4bit/raw/fa6f15978ba53de9ff3a95ce2821deb0ca4a15f0/model.safetensors.index.json" } ] }, { "id": "lmstudio-community/gemma-4-E4B-it-MLX-8bit", "name": "gemma-4-E4B-it-MLX-8bit 8-bit", "short_name": "gemma 4 E4B it MLX 8bit", "license": "apache-2.0", "hf_pipeline_tag": "any-to-any", "hf_downloads": 1632405, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "mlx", "any-to-any", "base_model:google/gemma-4-E4B-it", "base_model:quantized:google/gemma-4-E4B-it", "license:apache-2.0", "endpoints_compatible", "8-bit" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 7.996157418, "active_params_b": 7.996157418, "layers": 42, "attention_heads": 8, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "MLX 8-bit affine", "weight_bytes_per_param": 1.1179276805978458, "resident_weight_gb": 8.939125716, "active_weight_gb": 4.993745492, "kv_alloc_gb_per_1k_tokens": 0.086016, "kv_read_gb_per_1k_tokens": 0.086016, "notes": "Manual correction on 2026-07-02 from current Hugging Face API metadata, served MLX config, base Google config comparison, model card, safetensors index, and direct shard header range reads. Bound weights separate the exact 8.939125716 GB resident MLX package from the 4.993745492 GB ordinary text-decode swept traffic; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/gemma-4-E4B-it-MLX-8bit" }, { "label": "Model config", "url": "https://huggingface.co/lmstudio-community/gemma-4-E4B-it-MLX-8bit/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-E4B-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-E4B-it-MLX-8bit" }, { "label": "Safetensors index", "url": "https://huggingface.co/lmstudio-community/gemma-4-E4B-it-MLX-8bit/raw/c63b2f9519d800e591cb331e5c19c021f66bf79a/model.safetensors.index.json" } ] }, { "id": "lmstudio-community/gemma-4-E4B-it-MLX-5bit", "name": "gemma-4-E4B-it-MLX-5bit BF16/U32", "short_name": "gemma 4 E4B it MLX 5bit", "license": "apache-2.0", "hf_pipeline_tag": "any-to-any", "hf_downloads": 1623424, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "mlx", "any-to-any", "base_model:google/gemma-4-E4B-it", "base_model:quantized:google/gemma-4-E4B-it", "license:apache-2.0", "endpoints_compatible", "5-bit" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 7.996157418, "active_params_b": 7.996157418, "layers": 42, "attention_heads": 8, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "MLX 5-bit affine", "weight_bytes_per_param": 0.9200019563696888, "resident_weight_gb": 7.356480468, "active_weight_gb": 4.470276692, "kv_alloc_gb_per_1k_tokens": 0.086016, "kv_read_gb_per_1k_tokens": 0.086016, "notes": "Manual correction on 2026-07-02 from current Hugging Face API metadata, served MLX config, base Google config comparison, model card, safetensors index, and direct shard header range reads. Bound weights separate the exact 7.356480468 GB resident MLX package from the 4.470276692 GB ordinary text-decode swept traffic; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/gemma-4-E4B-it-MLX-5bit" }, { "label": "Model config", "url": "https://huggingface.co/lmstudio-community/gemma-4-E4B-it-MLX-5bit/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-E4B-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-E4B-it-MLX-5bit" }, { "label": "Safetensors index", "url": "https://huggingface.co/lmstudio-community/gemma-4-E4B-it-MLX-5bit/raw/cdb020657e3404e15a830bb61c6128dd77a2f6f3/model.safetensors.index.json" } ] }, { "id": "lmstudio-community/gemma-4-E4B-it-MLX-6bit", "name": "gemma-4-E4B-it-MLX-6bit BF16/U32", "short_name": "gemma 4 E4B it MLX 6bit", "license": "apache-2.0", "hf_pipeline_tag": "any-to-any", "hf_downloads": 1622196, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "mlx", "any-to-any", "base_model:google/gemma-4-E4B-it", "base_model:quantized:google/gemma-4-E4B-it", "license:apache-2.0", "endpoints_compatible", "6-bit", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 7.996157418, "active_params_b": 7.996157418, "layers": 42, "attention_heads": 8, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "MLX 6-bit affine", "weight_bytes_per_param": 0.9859771977790746, "resident_weight_gb": 7.884028884, "active_weight_gb": 4.644766292, "kv_alloc_gb_per_1k_tokens": 0.086016, "kv_read_gb_per_1k_tokens": 0.086016, "notes": "Manual correction on 2026-07-02 from current Hugging Face API metadata, served MLX config, base Google config comparison, model card, safetensors index, and direct shard header range reads. Bound weights separate the exact 7.884028884 GB resident MLX package from the 4.644766292 GB ordinary text-decode swept traffic; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/gemma-4-E4B-it-MLX-6bit" }, { "label": "Model config", "url": "https://huggingface.co/lmstudio-community/gemma-4-E4B-it-MLX-6bit/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-E4B-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-E4B-it-MLX-6bit" }, { "label": "Safetensors index", "url": "https://huggingface.co/lmstudio-community/gemma-4-E4B-it-MLX-6bit/raw/ff79ace56197c89ba21729e53c6378a96fe6367c/model.safetensors.index.json" } ] }, { "id": "moonshotai/Kimi-K2.5", "name": "Kimi-K2.5 4-bit", "short_name": "Kimi K2.5", "license": "other", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1520063, "tags": [ "transformers", "safetensors", "kimi_k25", "feature-extraction", "compressed-tensors", "image-text-to-text", "conversational", "custom_code", "arxiv:2602.02276", "license:other", "eval-results", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "kimi_k2 (DeepseekV3ForCausalLM)", "total_params_b": 1058.589420528, "active_params_b": 32, "layers": 61, "attention_heads": 64, "kv_heads": 64, "head_dim": 192, "max_context_tokens": 262144, "routed_experts": 384, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "INT4/BF16 compressed-tensors", "weight_bytes_per_param": 0.562208710190164, "resident_weight_gb": 595.148192736, "active_weight_gb": 32.986516736, "kv_alloc_gb_per_1k_tokens": 2.49856, "kv_read_gb_per_1k_tokens": 2.49856, "notes": "Manual correction on 2026-07-02 from current Hugging Face API metadata, served Kimi K2.5 config, model card, deployment guide, custom runtime review, safetensors index, and direct shard header range reads. Bound weights separate the exact 595.148192736 GB resident compressed-tensors package from 32.986516736 GB single-session ordinary text-decode MoE traffic; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/moonshotai/Kimi-K2.5" }, { "label": "Model config", "url": "https://huggingface.co/moonshotai/Kimi-K2.5/raw/main/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/moonshotai/Kimi-K2.5" }, { "label": "Safetensors index", "url": "https://huggingface.co/moonshotai/Kimi-K2.5/resolve/4d01dfe0332d63057c186e0b262165819efb6611/model.safetensors.index.json" } ] }, { "id": "meta-llama/Llama-3.1-8B", "name": "Llama-3.1-8B BF16", "short_name": "Llama 3.1 8B", "license": "llama3.1", "hf_pipeline_tag": "text-generation", "hf_downloads": 1517664, "tags": [ "transformers", "safetensors", "llama", "text-generation", "facebook", "meta", "pytorch", "llama-3", "en", "de", "fr", "it", "pt", "hi", "es", "th", "arxiv:2204.05149", "license:llama3.1", "eval-results", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gated config unavailable", "total_params_b": 8.030261248, "active_params_b": 8.030261248, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 16.060522496, "active_weight_gb": 16.060522496, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manual correction on 2026-07-02 from current Hugging Face API metadata and access checks. The API reports BF16 safetensors parameters, but raw config, README, safetensors index, and hf download with the configured osolmaz CLI token were denied because the repository requires approval. Production bounds use the fail-closed unsupported profile until config and tensor evidence are directly accessible." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/meta-llama/Llama-3.1-8B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.1-8B" }, { "label": "Gated config access check", "url": "https://huggingface.co/meta-llama/Llama-3.1-8B/raw/main/config.json" } ] }, { "id": "Qwen/Qwen2.5-Coder-32B-Instruct", "name": "Qwen2.5-Coder-32B-Instruct BF16", "short_name": "Qwen2.5 Coder 32B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 1445760, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "code", "codeqwen", "chat", "qwen", "qwen-coder", "conversational", "en", "arxiv:2409.12186", "arxiv:2309.00071", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-Coder-32B", "base_model:finetune:Qwen/Qwen2.5-Coder-32B", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 32.763876352, "active_params_b": 32.763876352, "layers": 64, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 65.527752704, "active_weight_gb": 63.970617344, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Manual correction on 2026-07-02 from current Hugging Face API metadata, model card, served config, base config comparison, safetensors index, and direct shard header range reads. Bound weights separate the exact 65.527752704 GB resident BF16 package from the 63.970617344 GB ordinary text-decode swept traffic; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-32B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Coder-32B-Instruct" }, { "label": "Safetensors index", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct/raw/381fc969f78efac66bc87ff7ddeadb7e73c218a7/model.safetensors.index.json" } ] }, { "id": "unsloth/gemma-4-26B-A4B-it-GGUF", "name": "gemma-4-26B-A4B-it-GGUF GGUF", "short_name": "gemma 4 26B A4B it", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1482262, "tags": [ "gguf", "gemma4", "unsloth", "gemma", "google", "image-text-to-text", "base_model:google/gemma-4-26B-A4B-it", "base_model:quantized:google/gemma-4-26B-A4B-it", "license:apache-2.0", "endpoints_compatible", "region:us", "imatrix", "conversational" ], "architecture": { "type": "MoE transformer", "detail": "gemma4_text", "total_params_b": 25.233142046, "active_params_b": 3.8, "layers": 30, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "GGUF MXFP4_MOE", "weight_bytes_per_param": 0.655924931, "resident_weight_gb": 16.551046944, "active_weight_gb": 3.450961856, "kv_alloc_gb_per_1k_tokens": 0.20992, "kv_read_gb_per_1k_tokens": 0.20992, "notes": "Manually audited on 2026-07-02 from HF API commit 3bb10d594514ef4edb7f3a65d41a7e4eb8c5767a, config.json, model card, linked GGUF file sizes, and a GGUF tensor-index range read. The authoritative bounds profile uses the selected gemma-4-26B-A4B-it-MXFP4_MOE.gguf artifact: 16.551046944 GB resident, 2.578677696 GB fixed traffic, 0.10903552 GB per routed expert, 8 routed experts per token, 128 routed experts, and layered Gemma 4 hybrid KV. Legacy per-1k KV fields show the short-context unsaturated coefficient; long-context calculation must use the profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/gemma-4-26B-A4B-it-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/unsloth/gemma-4-26B-A4B-it-GGUF/raw/3bb10d594514ef4edb7f3a65d41a7e4eb8c5767a/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/gemma-4-26B-A4B-it-GGUF" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/unsloth/gemma-4-26B-A4B-it-GGUF/resolve/3bb10d594514ef4edb7f3a65d41a7e4eb8c5767a/gemma-4-26B-A4B-it-MXFP4_MOE.gguf" } ] }, { "id": "Qwen/Qwen3.5-35B-A3B-FP8", "name": "Qwen3.5-35B-A3B-FP8 FP8", "short_name": "Qwen3.5 35B A3B FP8", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1403325, "tags": [ "transformers", "safetensors", "qwen3_5_moe", "image-text-to-text", "conversational", "base_model:Qwen/Qwen3.5-35B-A3B", "base_model:quantized:Qwen/Qwen3.5-35B-A3B", "license:apache-2.0", "endpoints_compatible", "fp8", "deploy:azure", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_5_moe_text", "total_params_b": 35.953925552, "active_params_b": 3, "layers": 40, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "FP8 F8_E4M3 + BF16 side tensors", "weight_bytes_per_param": 1, "resident_weight_gb": 37.454799072, "active_weight_gb": 3.481438336, "kv_alloc_gb_per_1k_tokens": 0.02048, "kv_read_gb_per_1k_tokens": 0.02048, "kv_alloc_fixed_gb": 0.06488064, "kv_read_fixed_gb": 0.06488064, "notes": "Manually audited on 2026-07-02 from HF API commit 9d1823d2dee688a6b25e77009dc727688c44936e, config.json, base config comparison, model card, and range-read safetensors headers. The authoritative bounds profile uses exact stored bytes: 37.454799072 GB resident, 2.474682496 GB fixed text traffic, 0.12584448 GB per routed expert, 8 routed experts plus 1 shared expert, 10 full-attention layers, and a fixed DeltaNet recurrent-state charge of 0.06488064 GB per session/token." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B-FP8" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B-FP8/raw/9d1823d2dee688a6b25e77009dc727688c44936e/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-35B-A3B-FP8" }, { "label": "Safetensors index", "url": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B-FP8/resolve/9d1823d2dee688a6b25e77009dc727688c44936e/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen2.5-VL-7B-Instruct-AWQ", "name": "Qwen2.5-VL-7B-Instruct-AWQ 4-bit", "short_name": "Qwen2.5 VL 7B Instruct", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1555903, "tags": [ "transformers", "safetensors", "qwen2_5_vl", "image-text-to-text", "multimodal", "conversational", "en", "arxiv:2309.00071", "arxiv:2409.12191", "arxiv:2308.12966", "base_model:Qwen/Qwen2.5-VL-7B-Instruct", "base_model:quantized:Qwen/Qwen2.5-VL-7B-Instruct", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "4-bit", "awq", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2_5_vl (Qwen2_5_VLForConditionalGeneration)", "total_params_b": 8.292166656, "active_params_b": 7.070619136, "layers": 28, "attention_heads": 28, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 128000, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "AWQ 4-bit + BF16/F16 side tensors", "weight_bytes_per_param": 0.5, "resident_weight_gb": 6.92384768, "active_weight_gb": 4.48075264, "kv_alloc_gb_per_1k_tokens": 0.057344, "kv_read_gb_per_1k_tokens": 0.057344, "notes": "Manually audited on 2026-07-02 from HF API commit 536a35794df8831aa814970ee8f89eff577e7718, config.json, base config comparison, model card, and range-read safetensors headers. The authoritative bounds profile uses exact stored bytes: 6.92384768 GB resident, 4.48075264 GB swept text traffic, 2.44309504 GB resident-only visual/input-embedding traffic, and full-context BF16 KV." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct-AWQ" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct-AWQ/raw/536a35794df8831aa814970ee8f89eff577e7718/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-VL-7B-Instruct-AWQ" }, { "label": "Safetensors index", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct-AWQ/raw/536a35794df8831aa814970ee8f89eff577e7718/model.safetensors.index.json" } ] }, { "id": "unsloth/gemma-4-12b-it-GGUF", "name": "Gemma 4 12B IT GGUF BF16", "short_name": "Gemma 4 12B IT GGUF", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1443180, "tags": [ "gguf", "gemma4", "unsloth", "gemma", "google", "gemma4_unified", "image-text-to-text", "base_model:google/gemma-4-12B-it", "base_model:quantized:google/gemma-4-12B-it", "license:apache-2.0", "endpoints_compatible", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_unified_text", "total_params_b": 11.907350576, "active_params_b": 11.907350576, "layers": 48, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16 GGUF", "weight_bytes_per_param": 2, "resident_weight_gb": 23.832065184, "active_weight_gb": 23.816242688, "kv_alloc_gb_per_1k_tokens": 0.335872, "kv_read_gb_per_1k_tokens": 0.335872, "notes": "Manually audited on 2026-07-02 from HF API commit 3249fa54d5efa384afc552cc6700ad091efd5c39, Google base config, model card, linked GGUF file sizes, and a GGUF tensor-index range read. The authoritative bounds profile uses the selected gemma-4-12b-it-BF16.gguf artifact: 23.832065184 GB resident, 23.816242688 GB swept text traffic, 0.015822496 GB resident-only GGUF metadata/header bytes, and layered Gemma 4 hybrid KV. Separate mmproj and MTP GGUF files are not included unless explicitly loaded. Legacy per-1k KV fields show the short-context unsaturated coefficient; long-context calculation must use the profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/gemma-4-12b-it-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/google/gemma-4-12B-it/raw/5926caa4ec0cac5cbfadaf4077420520de1d5205/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-12B-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/gemma-4-12b-it-GGUF" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/unsloth/gemma-4-12b-it-GGUF/resolve/3249fa54d5efa384afc552cc6700ad091efd5c39/gemma-4-12b-it-BF16.gguf" } ] }, { "id": "microsoft/Phi-3.5-vision-instruct", "name": "Phi-3.5-vision-instruct BF16", "short_name": "Phi 3.5 vision", "license": "mit", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1300939, "tags": [ "transformers", "safetensors", "phi3_v", "text-generation", "nlp", "code", "vision", "image-text-to-text", "conversational", "custom_code", "multilingual", "arxiv:2404.14219", "license:mit", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "phi3_v (Phi3VForCausalLM)", "total_params_b": 4.14662144, "active_params_b": 3.722578944, "layers": 32, "attention_heads": 32, "kv_heads": 32, "head_dim": 96, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 8.29324288, "active_weight_gb": 7.445157888, "kv_alloc_gb_per_1k_tokens": 0.393216, "kv_read_gb_per_1k_tokens": 0.393216, "notes": "Manually audited on 2026-07-02 from HF API commit 12b77fb40b63a2c73c68243d3f767aab688a1b2a, config.json, remote modeling code, model card, safetensors index, and range-read shard headers. The authoritative bounds profile uses exact BF16 stored bytes: 8.29324288 GB resident, 7.445157888 GB swept text/logit traffic, 0.848084992 GB resident-only vision plus input-embedding tensors, and full-context BF16 MHA KV." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/microsoft/Phi-3.5-vision-instruct" }, { "label": "Model config", "url": "https://huggingface.co/microsoft/Phi-3.5-vision-instruct/raw/12b77fb40b63a2c73c68243d3f767aab688a1b2a/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/microsoft/Phi-3.5-vision-instruct" }, { "label": "Safetensors index", "url": "https://huggingface.co/microsoft/Phi-3.5-vision-instruct/raw/12b77fb40b63a2c73c68243d3f767aab688a1b2a/model.safetensors.index.json" }, { "label": "Remote modeling code", "url": "https://huggingface.co/microsoft/Phi-3.5-vision-instruct/raw/12b77fb40b63a2c73c68243d3f767aab688a1b2a/modeling_phi3_v.py" } ] }, { "id": "QuantTrio/Qwen3-VL-30B-A3B-Instruct-AWQ", "name": "Qwen3-VL-30B-A3B-Instruct-AWQ 4-bit", "short_name": "Qwen3 VL 30B A3B Instruct", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 1398983, "tags": [ "transformers", "safetensors", "qwen3_vl_moe", "image-text-to-text", "AWQ", "vLLM", "text-generation", "conversational", "arxiv:2505.09388", "arxiv:2502.13923", "arxiv:2409.12191", "arxiv:2308.12966", "base_model:Qwen/Qwen3-VL-30B-A3B-Instruct", "base_model:quantized:Qwen/Qwen3-VL-30B-A3B-Instruct", "license:apache-2.0", "endpoints_compatible", "4-bit", "awq", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_vl_moe_text", "total_params_b": 31.070754032, "active_params_b": 3, "layers": 48, "attention_heads": 32, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "AWQ 4-bit + F16 side tensors", "weight_bytes_per_param": 0.5, "resident_weight_gb": 17.879935456, "active_weight_gb": 2.059956224, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "notes": "Audited from Hugging Face metadata on 2026-07-02; downloads=1398983; data quality: config, base config, API params, safetensors headers. Bound weights use exact stored AWQ I32/F16 tensor byte groups." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/QuantTrio/Qwen3-VL-30B-A3B-Instruct-AWQ" }, { "label": "Model config", "url": "https://huggingface.co/QuantTrio/Qwen3-VL-30B-A3B-Instruct-AWQ/raw/a5ea10732e2c9330957864ece3ec66d806a4b00d/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/QuantTrio/Qwen3-VL-30B-A3B-Instruct-AWQ" }, { "label": "Safetensors index", "url": "https://huggingface.co/QuantTrio/Qwen3-VL-30B-A3B-Instruct-AWQ/resolve/a5ea10732e2c9330957864ece3ec66d806a4b00d/model.safetensors.index.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct/raw/9c4b90e1e4ba969fd3b5378b57d966d725f1b86c/config.json" } ] }, { "id": "Qwen/Qwen2.5-1.5B", "name": "Qwen2.5-1.5B BF16", "short_name": "Qwen2.5 1.5B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 1209897, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "conversational", "en", "arxiv:2407.10671", "license:apache-2.0", "eval-results", "text-generation-inference", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 1.543714304, "active_params_b": 1.543714304, "layers": 28, "attention_heads": 12, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 3.087428608, "active_weight_gb": 3.087428608, "kv_alloc_gb_per_1k_tokens": 0.028672, "kv_read_gb_per_1k_tokens": 0.028672, "notes": "Audited from Hugging Face metadata on 2026-07-02; downloads=1209897; data quality: config, API params, safetensors header. Bound weights use exact BF16 tensor payload." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B/raw/8faed761d45a263340a0528343f099c05c9a4323/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-1.5B" }, { "label": "Safetensors header", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B/resolve/8faed761d45a263340a0528343f099c05c9a4323/model.safetensors" } ] }, { "id": "Qwen/Qwen3.5-122B-A10B-FP8", "name": "Qwen3.5-122B-A10B-FP8 FP8", "short_name": "Qwen3.5 122B A10B FP8", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1268604, "tags": [ "transformers", "safetensors", "qwen3_5_moe", "image-text-to-text", "conversational", "base_model:Qwen/Qwen3.5-122B-A10B", "base_model:quantized:Qwen/Qwen3.5-122B-A10B", "license:apache-2.0", "endpoints_compatible", "fp8", "deploy:azure", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_5_moe_text", "total_params_b": 125.094006512, "active_params_b": 10, "layers": 48, "attention_heads": 32, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "FP8 F8_E4M3 + BF16 side tensors", "weight_bytes_per_param": 1, "resident_weight_gb": 127.152313312, "active_weight_gb": 9.82643712, "kv_alloc_gb_per_1k_tokens": 0.024576, "kv_read_gb_per_1k_tokens": 0.024576, "kv_alloc_fixed_gb": 0.154533888, "kv_read_fixed_gb": 0.154533888, "notes": "Manually audited on 2026-07-02 from HF API commit a099dee70ccfcd8d5dda56aaa0b60cb8ecadabc9, config.json, base config comparison, and range-read safetensors headers. The authoritative bounds profile uses exact stored bytes: 127.152313312 GB resident, 6.202116096 GB fixed text traffic, 0.453040128 GB per routed expert, 8 routed experts plus 1 shared expert, 12 full-attention layers, and a fixed DeltaNet recurrent-state charge of 0.154533888 GB per session/token." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B-FP8" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B-FP8/raw/a099dee70ccfcd8d5dda56aaa0b60cb8ecadabc9/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B/raw/dc4d348443bc740c68e2d77492492c11606384d5/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-122B-A10B-FP8" }, { "label": "Safetensors index", "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B-FP8/resolve/a099dee70ccfcd8d5dda56aaa0b60cb8ecadabc9/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen2.5-14B-Instruct-AWQ", "name": "Qwen2.5-14B-Instruct-AWQ 4-bit", "short_name": "Qwen2.5 14B Instruct", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 1446368, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "chat", "conversational", "en", "arxiv:2309.00071", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-14B-Instruct", "base_model:quantized:Qwen/Qwen2.5-14B-Instruct", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "4-bit", "awq", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 14.770033664, "active_params_b": 14.770033664, "layers": 48, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "AWQ 4-bit + F16 side tensors", "weight_bytes_per_param": 0.5, "resident_weight_gb": 9.980028928, "active_weight_gb": 8.422893568, "kv_alloc_gb_per_1k_tokens": 0.196608, "kv_read_gb_per_1k_tokens": 0.196608, "notes": "Audited from Hugging Face metadata on 2026-07-02; downloads=1446368; data quality: config, base config, API params, safetensors headers. Bound weights use exact stored AWQ I32/F16 tensor byte groups." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-14B-Instruct-AWQ" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-14B-Instruct-AWQ/raw/539535859b135b0244c91f3e59816150c8056698/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-14B-Instruct" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen2.5-14B-Instruct/raw/cf98f3b3bbb457ad9e2bb7baf9a0125b6b88caa8/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-14B-Instruct-AWQ" }, { "label": "Safetensors index", "url": "https://huggingface.co/Qwen/Qwen2.5-14B-Instruct-AWQ/raw/539535859b135b0244c91f3e59816150c8056698/model.safetensors.index.json" } ] }, { "id": "nvidia/Gemma-4-31B-IT-NVFP4", "name": "Gemma-4-31B-IT-NVFP4 NVFP4", "short_name": "Gemma 4 31B IT NVFP4", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 1334156, "tags": [ "Model Optimizer", "safetensors", "nvidia", "ModelOpt", "Gemma-4-31B-IT", "lighthouse", "quantized", "NVFP4", "text-generation", "base_model:google/gemma-4-31B-it", "base_model:finetune:google/gemma-4-31B-it", "license:other", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text (Gemma4ForConditionalGeneration)", "total_params_b": 32.682372656, "active_params_b": 32.682372656, "layers": 60, "attention_heads": 32, "kv_heads": 16, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "NVFP4 ModelOpt + BF16/F8/U8 side tensors", "weight_bytes_per_param": 0.5, "resident_weight_gb": 32.633255032, "active_weight_gb": 31.48176796, "kv_alloc_gb_per_1k_tokens": 0.41984, "kv_read_gb_per_1k_tokens": 0.41984, "notes": "Manually audited on 2026-07-02 from HF API commit e5ef03afa233c35cb000323ff098d4291e1dd07c, config.json, hf_quant_config.json, base config comparison, and range-read safetensors headers. The authoritative bounds profile uses exact stored bytes: 32.633255032 GB resident, 31.48176796 GB swept text traffic, 1.151487072 GB resident-only vision traffic, and layered FP8 KV with 10 full-context layers plus 50 sliding-window layers." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/Gemma-4-31B-IT-NVFP4" }, { "label": "Model config", "url": "https://huggingface.co/nvidia/Gemma-4-31B-IT-NVFP4/raw/e5ef03afa233c35cb000323ff098d4291e1dd07c/config.json" }, { "label": "Quantization config", "url": "https://huggingface.co/nvidia/Gemma-4-31B-IT-NVFP4/raw/e5ef03afa233c35cb000323ff098d4291e1dd07c/hf_quant_config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-31B-it" }, { "label": "Base model config", "url": "https://huggingface.co/google/gemma-4-31B-it/raw/3548789868c5356dbf307c98e6f609007b82b3eb/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/Gemma-4-31B-IT-NVFP4" }, { "label": "Safetensors index", "url": "https://huggingface.co/nvidia/Gemma-4-31B-IT-NVFP4/raw/e5ef03afa233c35cb000323ff098d4291e1dd07c/model.safetensors.index.json" } ] }, { "id": "meta-llama/Meta-Llama-3-8B-Instruct", "name": "Meta-Llama-3-8B-Instruct BF16", "short_name": "Meta Llama 3 8B", "license": "llama3", "hf_pipeline_tag": "text-generation", "hf_downloads": 1306206, "tags": [ "transformers", "safetensors", "llama", "text-generation", "facebook", "meta", "pytorch", "llama-3", "conversational", "en", "license:llama3", "eval-results", "text-generation-inference", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gated config unavailable", "total_params_b": 8.030261248, "active_params_b": 8.030261248, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 16.060522496, "active_weight_gb": 16.060522496, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manual correction on 2026-07-02 from current Hugging Face API metadata and access checks. The API reports BF16 safetensors parameters, but raw config, README, safetensors index, and hf download with the configured CLI token were denied because the repository requires approval. Production bounds use the fail-closed unsupported profile until config and tensor evidence are directly accessible." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/meta-llama/Meta-Llama-3-8B-Instruct" }, { "label": "Gated config access check", "url": "https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct/raw/8afb486c1db24fe5011ec46dfbe5b5dccdb575c2/config.json" } ] }, { "id": "google/diffusiongemma-26B-A4B-it", "name": "diffusiongemma-26B-A4B-it BF16", "short_name": "diffusiongemma 26B A4B it", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1424776, "tags": [ "transformers", "safetensors", "diffusion_gemma", "image-text-to-text", "conversational", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "diffusion_gemma_text (DiffusionGemmaForBlockDiffusion)", "total_params_b": 25.823781228, "active_params_b": 4, "layers": 30, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 51.647562456, "active_weight_gb": 7.680751164, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manual correction on 2026-07-02 from HF API commit 0f28bc42f588fbd8f71e08102b1c3960298a1358, config.json, model card, and range-read safetensors headers. The exact BF16 resident package is 51.647562456 GB. The profile is fail-closed because DiffusionGemma uses block diffusion over a 256-token canvas, which Bounds Engine v1 does not model as ordinary autoregressive decode." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/diffusiongemma-26B-A4B-it" }, { "label": "Model config", "url": "https://huggingface.co/google/diffusiongemma-26B-A4B-it/raw/0f28bc42f588fbd8f71e08102b1c3960298a1358/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/diffusiongemma-26B-A4B-it" }, { "label": "Safetensors index", "url": "https://huggingface.co/google/diffusiongemma-26B-A4B-it/raw/0f28bc42f588fbd8f71e08102b1c3960298a1358/model.safetensors.index.json" } ] }, { "id": "nvidia/Kimi-K2.5-NVFP4", "name": "Kimi-K2.5-NVFP4 NVFP4", "short_name": "Kimi K2.5 NVFP4", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 1206834, "tags": [ "Model Optimizer", "safetensors", "kimi_k25", "nvidia", "ModelOpt", "Kimi-K2", "quantized", "FP4", "text-generation", "conversational", "custom_code", "base_model:moonshotai/Kimi-K2.5", "base_model:quantized:moonshotai/Kimi-K2.5", "license:other", "modelopt", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "kimi_k25 (KimiK25ForConditionalGeneration / DeepseekV3ForCausalLM)", "total_params_b": 1058.589420528, "active_params_b": 32, "layers": 61, "attention_heads": 64, "kv_heads": 64, "head_dim": 192, "max_context_tokens": 262144, "routed_experts": 384, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "NVFP4/BF16/FP8 ModelOpt", "weight_bytes_per_param": 0.5, "resident_weight_gb": 590.779912064, "active_weight_gb": 28.618236064, "kv_alloc_gb_per_1k_tokens": 1.24928, "kv_read_gb_per_1k_tokens": 1.24928, "notes": "Manual correction on 2026-07-02 from current Hugging Face API metadata, NVIDIA model card, served config, hf_quant_config, custom runtime review, safetensors index, and direct shard header range reads. Bound weights separate the exact 590.779912064 GB resident ModelOpt package from 28.618236064 GB single-session ordinary text-decode MoE traffic; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/Kimi-K2.5-NVFP4" }, { "label": "Model config", "url": "https://huggingface.co/nvidia/Kimi-K2.5-NVFP4/raw/0fd0a5e6879298d3476e3b61852a79792a35ae3d/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/Kimi-K2.5-NVFP4" }, { "label": "ModelOpt quantization config", "url": "https://huggingface.co/nvidia/Kimi-K2.5-NVFP4/raw/0fd0a5e6879298d3476e3b61852a79792a35ae3d/hf_quant_config.json" }, { "label": "Safetensors index", "url": "https://huggingface.co/nvidia/Kimi-K2.5-NVFP4/resolve/0fd0a5e6879298d3476e3b61852a79792a35ae3d/model.safetensors.index.json" }, { "label": "Base model card", "url": "https://huggingface.co/moonshotai/Kimi-K2.5" } ] }, { "id": "Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8", "name": "Qwen3-Coder-30B-A3B-Instruct-FP8 FP8", "short_name": "Qwen3 Coder 30B A3B Instruct FP8", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 1358940, "tags": [ "transformers", "safetensors", "qwen3_moe", "text-generation", "conversational", "arxiv:2505.09388", "license:apache-2.0", "endpoints_compatible", "fp8", "deploy:azure", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_moe (Qwen3MoeForCausalLM)", "total_params_b": 30.533947392, "active_params_b": 3.042033664, "layers": 48, "attention_heads": 32, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "FP8/BF16", "weight_bytes_per_param": 1, "resident_weight_gb": 31.170895872, "active_weight_gb": 3.366158336, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "notes": "Manual correction on 2026-07-02 from current Hugging Face API metadata, model card, served config, BF16 config comparison, safetensors index, and direct safetensors header range reads. Active traffic is fixed decode traffic plus 8 routed expert groups; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8/raw/dcaee4d4dfc5ee71ad501f01f530e5652438fde0/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8" }, { "label": "Safetensors index", "url": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8/resolve/dcaee4d4dfc5ee71ad501f01f530e5652438fde0/model.safetensors.index.json" }, { "label": "BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct/raw/b2cff646eb4bb1d68355c01b18ae02e7cf42d120/config.json" } ] }, { "id": "Qwen/Qwen3-Coder-Next", "name": "Qwen3-Coder-Next BF16", "short_name": "Qwen3 Coder Next", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 1272766, "tags": [ "transformers", "safetensors", "qwen3_next", "text-generation", "conversational", "license:apache-2.0", "eval-results", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_next (Qwen3NextForCausalLM)", "total_params_b": 79.674391296, "active_params_b": 3.56376448, "layers": 48, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 159.348782592, "active_weight_gb": 7.12752896, "kv_alloc_gb_per_1k_tokens": 0.024576, "kv_read_gb_per_1k_tokens": 0.024576, "kv_alloc_fixed_gb": 0.077856768, "kv_read_fixed_gb": 0.077856768, "notes": "Manual correction on 2026-07-02 from current Hugging Face API metadata, model card, served config, safetensors index, direct safetensors header range reads, and Qwen3-Next runtime review. Active traffic is fixed decode traffic plus 10 routed expert groups; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-Coder-Next" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-Coder-Next/raw/a7fbcb5c0e12d62a448eaa0e260346bf5dcc0feb/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-Coder-Next" }, { "label": "Safetensors index", "url": "https://huggingface.co/Qwen/Qwen3-Coder-Next/resolve/a7fbcb5c0e12d62a448eaa0e260346bf5dcc0feb/model.safetensors.index.json" }, { "label": "Transformers Qwen3-Next implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/qwen3_next/modeling_qwen3_next.py" } ] }, { "id": "Qwen/Qwen3.5-27B-FP8", "name": "Qwen3.5-27B-FP8 FP8", "short_name": "Qwen3.5 27B FP8", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1155275, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "conversational", "base_model:Qwen/Qwen3.5-27B", "base_model:quantized:Qwen/Qwen3.5-27B", "license:apache-2.0", "endpoints_compatible", "fp8", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 27.782935472, "active_params_b": 25.626084864, "layers": 64, "attention_heads": 24, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "FP8/BF16/F32", "weight_bytes_per_param": 1, "resident_weight_gb": 30.86668016, "active_weight_gb": 26.925223424, "kv_alloc_gb_per_1k_tokens": 0.065536, "kv_read_gb_per_1k_tokens": 0.065536, "kv_alloc_fixed_gb": 0.154927104, "kv_read_fixed_gb": 0.154927104, "notes": "Manual correction on 2026-07-02 from current Hugging Face API metadata, model card, served config, base config comparison, safetensors index, direct safetensors header range reads, and Qwen3.5 runtime review. Active traffic is ordinary text swept traffic excluding visual/MTP/input-embedding resident bytes; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3.5-27B-FP8" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3.5-27B-FP8/raw/97f5941bf617e31c5e237364a8602ce3f03a551a/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-27B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-27B-FP8" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3.5-27B/raw/fc05daec18b0a78c049392ed2e771dde82bdf654/config.json" }, { "label": "Safetensors index", "url": "https://huggingface.co/Qwen/Qwen3.5-27B-FP8/resolve/97f5941bf617e31c5e237364a8602ce3f03a551a/model.safetensors.index.json" }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py" } ] }, { "id": "meta-llama/Meta-Llama-3-8B", "name": "Meta-Llama-3-8B BF16", "short_name": "Meta Llama 3 8B", "license": "llama3", "hf_pipeline_tag": "text-generation", "hf_downloads": 1237078, "tags": [ "transformers", "safetensors", "llama", "text-generation", "facebook", "meta", "pytorch", "llama-3", "en", "license:llama3", "eval-results", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gated config unavailable", "total_params_b": 8.030261248, "active_params_b": 8.030261248, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 16.060522496, "active_weight_gb": 16.060522496, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manual correction on 2026-07-02 from current Hugging Face API metadata and access checks. The API reports BF16 safetensors parameters, but raw config, README, safetensors index, and hf download with the configured CLI token were denied because the repository requires approval. Production bounds use the fail-closed unsupported profile until config and tensor evidence are directly accessible." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/meta-llama/Meta-Llama-3-8B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/meta-llama/Meta-Llama-3-8B" }, { "label": "Gated config access check", "url": "https://huggingface.co/meta-llama/Meta-Llama-3-8B/raw/8cde5ca8380496c9a6cc7ef3a8b46a0372a1d920/config.json" } ] }, { "id": "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16", "name": "NVIDIA-Nemotron-3-Super-120B-A12B-BF16 F32/BF16", "short_name": "NVIDIA Nemotron 3 Super 120B A12B BF16", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 1260453, "tags": [ "transformers", "safetensors", "nemotron_h", "text-generation", "nvidia", "pytorch", "nemotron-3", "latent-moe", "mtp", "conversational", "custom_code", "en", "fr", "es", "it", "de", "ja", "zh", "dataset:nvidia/nemotron-post-training-v3", "dataset:nvidia/nemotron-pre-training-datasets", "arxiv:2512.20848", "arxiv:2512.20856", "license:other", "eval-results", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "nemotron_h (NemotronHForCausalLM)", "total_params_b": 123.611033088, "active_params_b": 12.233387008, "layers": 88, "attention_heads": 32, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 512, "routed_experts_per_token": 22, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "F32/BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 247.22210816, "active_weight_gb": 24.466774016, "kv_alloc_gb_per_1k_tokens": 0.004096, "kv_read_gb_per_1k_tokens": 0.004096, "kv_alloc_fixed_gb": 0.17104896, "kv_read_fixed_gb": 0.17104896, "expert_param_b": 0.22020096, "fixed_param_b": 7.388965888, "shared_expert_param_b": 0, "notes": "Manual correction on 2026-07-02 from current Hugging Face API metadata, pinned served config, BF16 base config comparison, custom runtime review, safetensors index, and direct range-read safetensors shard headers. Legacy fallback values mirror the audited profile: exact stored resident bytes, fixed always-on hybrid/shared-expert traffic, expected distinct routed experts, FP8 attention KV, and BF16/F32 Mamba recurrent state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16" }, { "label": "Pinned served config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16/raw/d51eab0d1f979ebc26b546e634a04f450d99158e/config.json" }, { "label": "Pinned base config comparison", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-Base-BF16/raw/46cc6113d364942e7742b0b2afd35b5db5058b29/config.json" }, { "label": "Custom runtime code", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16/raw/d51eab0d1f979ebc26b546e634a04f450d99158e/modeling_nemotron_h.py" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16/resolve/d51eab0d1f979ebc26b546e634a04f450d99158e/model.safetensors.index.json" } ] }, { "id": "nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1", "name": "Llama-3.1-Nemotron-Nano-VL-8B-V1 BF16", "short_name": "Llama 3.1 Nemotron Nano VL 8B V1", "license": "other", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1220336, "tags": [ "transformers", "safetensors", "nvidia", "VLM", "llama3.1", "image-text-to-text", "license:other", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "Llama_Nemotron_Nano_VL (Llama 3.1 + C-RADIOv2-H)", "total_params_b": 8.721768454, "active_params_b": 7.505973248, "layers": 32, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 16384, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 17.443536908, "active_weight_gb": 15.011946496, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "notes": "Manual correction on 2026-07-02 from current Hugging Face API metadata, model card, resolved pinned config, custom runtime code, image processor code, and direct range-read safetensors header evidence. Bound weights separate the 17.443536908 GB resident multimodal package from the 15.011946496 GB ordinary text-decode swept subset; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1" }, { "label": "Resolved pinned config", "url": "https://huggingface.co/nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1/resolve/437f4e28b989cc2d9a16b6767cc930cdf48797ff/config.json" }, { "label": "Custom runtime code", "url": "https://huggingface.co/nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1/resolve/437f4e28b989cc2d9a16b6767cc930cdf48797ff/modeling.py" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1/resolve/437f4e28b989cc2d9a16b6767cc930cdf48797ff/model.safetensors" } ] }, { "id": "QuantTrio/Qwen3.6-27B-AWQ", "name": "Qwen3.6-27B-AWQ 4-bit", "short_name": "Qwen3.6 27B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1353250, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "vLLM", "AWQ", "conversational", "base_model:Qwen/Qwen3.6-27B", "base_model:quantized:Qwen/Qwen3.6-27B", "license:apache-2.0", "endpoints_compatible", "4-bit", "awq", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 27.781427952, "active_params_b": 25.624600064, "layers": 64, "attention_heads": 24, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "4-bit", "weight_bytes_per_param": 0.5, "resident_weight_gb": 21.852837344, "active_weight_gb": 17.539181568, "kv_alloc_gb_per_1k_tokens": 0.065536, "kv_read_gb_per_1k_tokens": 0.065536, "kv_alloc_fixed_gb": 0.154927104, "kv_read_fixed_gb": 0.154927104, "notes": "Manual correction on 2026-07-02 from current Hugging Face API metadata, model card, pinned AWQ config, current base config comparison, safetensors index, and direct range-read shard headers. Bound weights separate the 21.852837344 GB resident AWQ package from the 17.539181568 GB ordinary text-decode swept subset; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/QuantTrio/Qwen3.6-27B-AWQ" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/QuantTrio/Qwen3.6-27B-AWQ" }, { "label": "Pinned AWQ config", "url": "https://huggingface.co/QuantTrio/Qwen3.6-27B-AWQ/raw/9b507bdc9afafb87b7898700cc2a591aa6639461/config.json" }, { "label": "Base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/QuantTrio/Qwen3.6-27B-AWQ/resolve/9b507bdc9afafb87b7898700cc2a591aa6639461/model.safetensors.index.json" } ] }, { "id": "deepseek-ai/DeepSeek-V4-Pro", "name": "DeepSeek-V4-Pro FP4/FP8", "short_name": "DeepSeek V4 Pro", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 1140384, "tags": [ "transformers", "safetensors", "deepseek_v4", "text-generation", "arxiv:2606.19348", "license:mit", "eval-results", "endpoints_compatible", "8-bit", "fp8", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "deepseek_v4 (DeepseekV4ForCausalLM)", "total_params_b": 1600, "active_params_b": 49, "layers": 61, "attention_heads": 128, "kv_heads": 1, "head_dim": 512, "max_context_tokens": 1048576, "routed_experts": 384, "routed_experts_per_token": 6, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "FP4 + FP8 mixed", "weight_bytes_per_param": 1, "resident_weight_gb": 864.704792696, "active_weight_gb": 39.6852203, "kv_alloc_gb_per_1k_tokens": 0.009848, "kv_read_gb_per_1k_tokens": 0.002168, "kv_alloc_fixed_gb": 0.02670592, "kv_read_fixed_gb": 0.039452672, "expert_param_b": 2.140766208, "shared_expert_param_b": 4.02992352, "fixed_param_b": 22.810699532, "notes": "Manual correction on 2026-07-02 from current Hugging Face API metadata, model card, pinned config, official inference config/code, safetensors index, and direct range-read shard headers. Bound weights use exact stored mixed FP4/FP8 shard bytes and the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro" }, { "label": "Model config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/raw/b5968e9190ef611bbf34a7229255be88a0e937c1/config.json" }, { "label": "Official inference config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/raw/b5968e9190ef611bbf34a7229255be88a0e937c1/inference/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/resolve/b5968e9190ef611bbf34a7229255be88a0e937c1/model.safetensors.index.json" } ] }, { "id": "deepseek-ai/DeepSeek-V3-0324", "name": "DeepSeek-V3-0324 FP8", "short_name": "DeepSeek V3 0324", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 1108710, "tags": [ "transformers", "safetensors", "deepseek_v3", "text-generation", "conversational", "custom_code", "arxiv:2412.19437", "license:mit", "eval-results", "text-generation-inference", "endpoints_compatible", "fp8", "region:us" ], "architecture": { "type": "MLA MoE transformer", "detail": "deepseek_v3 (DeepseekV3ForCausalLM)", "total_params_b": 671, "active_params_b": 37, "layers": 61, "attention_heads": 128, "kv_heads": 128, "head_dim": 128, "max_context_tokens": 163840, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "kv_lora_rank": 512, "qk_rope_head_dim": 64, "qk_nope_head_dim": 128, "v_head_dim": 128 }, "adapter": { "kind": "moe", "weight_precision": "FP8", "weight_bytes_per_param": 1, "resident_weight_gb": 688.57483936, "active_weight_gb": 39.521833312, "kv_alloc_gb_per_1k_tokens": 0.070272, "kv_read_gb_per_1k_tokens": 0.070272, "expert_param_b": 2.554954752, "shared_expert_param_b": 2.554954752, "fixed_param_b": 16.527240544, "notes": "Manual correction on 2026-07-02 from current Hugging Face API metadata, model card, pinned config, official DeepSeek-V3 local inference config/code, custom HF modeling caveat, safetensors index, and direct range-read shard headers. Bound weights use exact stored FP8/BF16/F32 shard bytes and MLA cache geometry; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3-0324" }, { "label": "Model config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3-0324/raw/e9b33add76883f293d6bf61f6bd89b497e80e335/config.json" }, { "label": "DeepSeek-V3 local inference code", "url": "https://github.com/deepseek-ai/DeepSeek-V3/blob/9b4e9788e4a3a731f7567338ed15d3ec549ce03b/inference/model.py" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3-0324/resolve/e9b33add76883f293d6bf61f6bd89b497e80e335/model.safetensors.index.json" } ] }, { "id": "nvidia/diffusiongemma-26B-A4B-it-NVFP4", "name": "diffusiongemma-26B-A4B-it-NVFP4 NVFP4", "short_name": "diffusiongemma 26B A4B it NVFP4", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 1307197, "tags": [ "safetensors", "diffusion_gemma", "nvidia", "ModelOpt", "DiffusionGemma-26B-A4B-IT", "quantized", "NVFP4", "nvfp4", "text-generation", "conversational", "base_model:google/diffusiongemma-26B-A4B-it", "base_model:quantized:google/diffusiongemma-26B-A4B-it", "license:apache-2.0", "8-bit", "modelopt", "region:us" ], "architecture": { "type": "Block diffusion MoE transformer", "detail": "diffusion_gemma (DiffusionGemmaForBlockDiffusion)", "total_params_b": 25.2, "active_params_b": 3.8, "layers": 30, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "NVFP4", "weight_bytes_per_param": 0.5, "resident_weight_gb": 18.818050776, "active_weight_gb": 5.628906684, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manual correction on 2026-07-02 from current HF API commit 2ea837236295d617ac27f8c17d61228081932c40, pinned config, hf_quant_config, generation config, model card, safetensors index, and direct range-read shard headers. The exact ModelOpt resident package is 18.818050776 GB. Production bounds are disabled by the unsupported profile because DiffusionGemma uses block diffusion over a 256-token canvas, which Bounds Engine v1 does not model as ordinary autoregressive decode." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/diffusiongemma-26B-A4B-it-NVFP4" }, { "label": "Model config", "url": "https://huggingface.co/nvidia/diffusiongemma-26B-A4B-it-NVFP4/raw/2ea837236295d617ac27f8c17d61228081932c40/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/diffusiongemma-26B-A4B-it" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/nvidia/diffusiongemma-26B-A4B-it-NVFP4/raw/2ea837236295d617ac27f8c17d61228081932c40/model.safetensors.index.json" } ] }, { "id": "deepseek-ai/DeepSeek-V3", "name": "DeepSeek-V3 FP8", "short_name": "DeepSeek V3", "license": "deepseek-model-license", "hf_pipeline_tag": "text-generation", "hf_downloads": 1113089, "tags": [ "transformers", "safetensors", "deepseek_v3", "text-generation", "conversational", "custom_code", "arxiv:2412.19437", "eval-results", "text-generation-inference", "endpoints_compatible", "fp8", "region:us" ], "architecture": { "type": "MLA MoE transformer", "detail": "deepseek_v3 (DeepseekV3ForCausalLM)", "total_params_b": 671, "active_params_b": 37, "layers": 61, "attention_heads": 128, "kv_heads": 128, "head_dim": 128, "max_context_tokens": 163840, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "kv_lora_rank": 512, "qk_rope_head_dim": 64, "qk_nope_head_dim": 128, "v_head_dim": 128 }, "adapter": { "kind": "moe", "weight_precision": "FP8", "weight_bytes_per_param": 1, "resident_weight_gb": 688.57483936, "active_weight_gb": 39.521833312, "kv_alloc_gb_per_1k_tokens": 0.070272, "kv_read_gb_per_1k_tokens": 0.070272, "expert_param_b": 2.554954752, "shared_expert_param_b": 2.554954752, "fixed_param_b": 16.527240544, "notes": "Manual correction on 2026-07-02 from current HF API metadata, pinned config, model card, official DeepSeek-V3 local inference config/code, custom HF modeling caveat, safetensors index, direct range-read shard headers, and model license. Bound weights use exact stored FP8/BF16/F32 shard bytes and official absorb-MLA cache geometry; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3" }, { "label": "Model config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3/raw/e815299b0bcbac849fa540c768ef21845365c9eb/config.json" }, { "label": "DeepSeek-V3 local inference code", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3/raw/e815299b0bcbac849fa540c768ef21845365c9eb/inference/model.py" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3/resolve/e815299b0bcbac849fa540c768ef21845365c9eb/model.safetensors.index.json" } ] }, { "id": "unsloth/Qwen3.6-27B-NVFP4", "name": "Qwen3.6-27B-NVFP4 NVFP4", "short_name": "Qwen3.6 27B NVFP4", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1149369, "tags": [ "safetensors", "qwen3_5", "unsloth", "image-text-to-text", "conversational", "base_model:Qwen/Qwen3.6-27B", "base_model:quantized:Qwen/Qwen3.6-27B", "license:apache-2.0", "8-bit", "compressed-tensors", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 27.781427952, "active_params_b": 25.624600064, "layers": 64, "attention_heads": 24, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "NVFP4", "weight_bytes_per_param": 0.5, "resident_weight_gb": 26.380988256, "active_weight_gb": 22.06733248, "kv_alloc_gb_per_1k_tokens": 0.065536, "kv_read_gb_per_1k_tokens": 0.065536, "kv_alloc_fixed_gb": 0.154927104, "kv_read_fixed_gb": 0.154927104, "notes": "Manual correction on 2026-07-02 from current Hugging Face API metadata, model card, pinned NVFP4 config and recipe, current base config comparison, and direct range-read safetensors header. Bound weights separate the 26.380988256 GB resident NVFP4/BF16/F8/F32 package from the 22.067332480 GB ordinary text-decode swept subset; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/Qwen3.6-27B-NVFP4" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3.6-27B-NVFP4" }, { "label": "Pinned NVFP4 config", "url": "https://huggingface.co/unsloth/Qwen3.6-27B-NVFP4/raw/890bdef7a42feba6d83b6e17a03315c694112f2a/config.json" }, { "label": "Pinned NVFP4 recipe", "url": "https://huggingface.co/unsloth/Qwen3.6-27B-NVFP4/raw/890bdef7a42feba6d83b6e17a03315c694112f2a/nvfp4_recipe.json" }, { "label": "Base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/unsloth/Qwen3.6-27B-NVFP4/resolve/890bdef7a42feba6d83b6e17a03315c694112f2a/model.safetensors" } ] }, { "id": "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct", "name": "DeepSeek-Coder-V2-Lite-Instruct BF16", "short_name": "DeepSeek Coder V2 Lite", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 1094936, "tags": [ "transformers", "safetensors", "deepseek_v2", "text-generation", "conversational", "custom_code", "arxiv:2401.06066", "license:other", "eval-results", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "deepseek_v2 (DeepseekV2ForCausalLM)", "total_params_b": 15.706484224, "active_params_b": 2.451435008, "layers": 27, "attention_heads": 16, "kv_heads": 16, "head_dim": 128, "max_context_tokens": 163840, "routed_experts": 64, "routed_experts_per_token": 6, "shared_experts_per_token": 2 }, "adapter": { "kind": "moe", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 31.412968448, "active_weight_gb": 4.902870016, "kv_alloc_gb_per_1k_tokens": 0.27648, "kv_read_gb_per_1k_tokens": 0.27648, "notes": "Manual correction on 2026-07-02 from current Hugging Face API metadata, model card, pinned config, pinned custom modeling code, safetensors index, and direct range-read shard headers. Bound weights separate the 31.412968448 GB resident BF16 checkpoint from 4.902870016 GB single-token active ordinary decode traffic; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct" }, { "label": "Pinned model config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct/raw/e434a23f91ba5b4923cf6c9d9a238eb4a08e3a11/config.json" }, { "label": "Pinned custom modeling code", "url": "https://huggingface.co/deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct/raw/e434a23f91ba5b4923cf6c9d9a238eb4a08e3a11/modeling_deepseek.py" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct/resolve/e434a23f91ba5b4923cf6c9d9a238eb4a08e3a11/model.safetensors.index.json" } ] }, { "id": "microsoft/Phi-3.5-mini-instruct", "name": "Phi-3.5-mini-instruct BF16", "short_name": "Phi 3.5 mini", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 1030083, "tags": [ "transformers", "safetensors", "phi3", "text-generation", "nlp", "code", "conversational", "custom_code", "multilingual", "arxiv:2404.14219", "arxiv:2407.13833", "arxiv:2403.06412", "license:mit", "eval-results", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "phi3 (Phi3ForCausalLM)", "total_params_b": 3.821079552, "active_params_b": 3.722578944, "layers": 32, "attention_heads": 32, "kv_heads": 32, "head_dim": 96, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 7.642159104, "active_weight_gb": 7.445157888, "kv_alloc_gb_per_1k_tokens": 0.393216, "kv_read_gb_per_1k_tokens": 0.393216, "notes": "Manual correction on 2026-07-02 from current Hugging Face API metadata, model card, pinned config, pinned remote modeling code, safetensors index, and direct range-read shard headers. Bound weights separate the 7.642159104 GB resident BF16 checkpoint from the 7.445157888 GB ordinary text-decode swept subset; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/microsoft/Phi-3.5-mini-instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/microsoft/Phi-3.5-mini-instruct" }, { "label": "Pinned model config", "url": "https://huggingface.co/microsoft/Phi-3.5-mini-instruct/raw/2fe192450127e6a83f7441aef6e3ca586c338b77/config.json" }, { "label": "Pinned remote modeling code", "url": "https://huggingface.co/microsoft/Phi-3.5-mini-instruct/raw/2fe192450127e6a83f7441aef6e3ca586c338b77/modeling_phi3.py" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/microsoft/Phi-3.5-mini-instruct/resolve/2fe192450127e6a83f7441aef6e3ca586c338b77/model.safetensors.index.json" } ] }, { "id": "lmstudio-community/gemma-4-26B-A4B-it-QAT-MLX-4bit", "name": "gemma-4-26B-A4B-it-QAT-MLX-4bit 4-bit", "short_name": "gemma 4 26B A4B it QAT MLX 4bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1191518, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "mlx", "conversational", "license:apache-2.0", "endpoints_compatible", "4-bit", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "gemma4_text", "total_params_b": 26.544131376, "active_params_b": 4, "layers": 30, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "MLX QAT 4-bit affine", "weight_bytes_per_param": 0.5880250448923184, "resident_weight_gb": 15.608614044, "active_weight_gb": 2.424219708, "kv_alloc_gb_per_1k_tokens": 0.20992, "kv_read_gb_per_1k_tokens": 0.20992, "notes": "Manual correction on 2026-07-02 from current Hugging Face API metadata, served MLX config, public Google QAT config comparison, non-QAT Google config comparison, safetensors index, and direct range-read shard headers. Bound weights separate the exact 15.608614044 GB resident MLX package from 1.621321788 GB fixed ordinary text traffic plus 0.100362240 GB per routed expert index; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-QAT-MLX-4bit" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-26B-A4B-it-QAT-MLX-4bit" }, { "label": "Pinned MLX model config", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-QAT-MLX-4bit/raw/f03a4a76828804e3b56758627383fbc7bd32015c/config.json" }, { "label": "Public Google QAT reference config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it-qat-q4_0-unquantized/raw/641f184470aa8554ae7957599a624badc2bf4e57/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-QAT-MLX-4bit/raw/f03a4a76828804e3b56758627383fbc7bd32015c/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen2.5-Coder-14B-Instruct-AWQ", "name": "Qwen2.5-Coder-14B-Instruct-AWQ 4-bit", "short_name": "Qwen2.5 Coder 14B Instruct", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 1077947, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "code", "codeqwen", "chat", "qwen", "qwen-coder", "conversational", "en", "arxiv:2409.12186", "arxiv:2309.00071", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-Coder-14B-Instruct", "base_model:quantized:Qwen/Qwen2.5-Coder-14B-Instruct", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "4-bit", "awq", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 14.770033664, "active_params_b": 13.991465984, "layers": 48, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "4-bit", "weight_bytes_per_param": 0.5, "resident_weight_gb": 9.980028928, "active_weight_gb": 8.422893568, "kv_alloc_gb_per_1k_tokens": 0.196608, "kv_read_gb_per_1k_tokens": 0.196608, "notes": "Manual correction on 2026-07-02 from current Hugging Face API metadata, model card, served AWQ config, audited Qwen2.5-Coder 14B Instruct base profile comparison, safetensors index, and direct safetensors header range reads. Bound weights separate the 9.980028928 GB resident AWQ package from the 8.422893568 GB ordinary text-decode swept subset; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-14B-Instruct-AWQ" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Coder-14B-Instruct-AWQ" }, { "label": "Pinned AWQ config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-14B-Instruct-AWQ/raw/eb3172f06a6d6b3a15f08947b0668d782e4d2d2c/config.json" }, { "label": "Audited BF16 base profile source", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-14B-Instruct" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-14B-Instruct-AWQ/raw/eb3172f06a6d6b3a15f08947b0668d782e4d2d2c/model.safetensors.index.json" } ] }, { "id": "RedHatAI/gemma-4-26B-A4B-it-NVFP4", "name": "gemma-4-26B-A4B-it-NVFP4 NVFP4", "short_name": "gemma 4 26B A4B it NVFP4", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1090344, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "fp4", "vllm", "llm-compressor", "compressed-tensors", "conversational", "base_model:google/gemma-4-26B-A4B-it", "base_model:quantized:google/gemma-4-26B-A4B-it", "license:apache-2.0", "endpoints_compatible", "8-bit", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "gemma4_text", "total_params_b": 26.544131376, "active_params_b": 4, "layers": 30, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "NVFP4", "weight_bytes_per_param": 0.5, "resident_weight_gb": 16.417035012, "active_weight_gb": 3.22789098, "kv_alloc_gb_per_1k_tokens": 0.20992, "kv_read_gb_per_1k_tokens": 0.20992, "notes": "Manually audited on 2026-07-02 from HF API commit 24eef434bc1089b25f8866c687fd015c63546735, config.json, README/model card, Google base config comparison, safetensors index, and a safetensors header range read. The authoritative bounds profile uses exact stored bytes for the RedHatAI NVFP4 artifact: 16.417035012 GB resident, 2.4249873 GB fixed text-decode traffic, 0.10036296 GB per routed expert, 8 routed experts per token, 128 routed experts, 1 shared expert, tied embedding/logit projection charged once, and layered Gemma 4 BF16 KV because kv_cache_scheme is null. Legacy per-1k KV fields show the short-context unsaturated coefficient; long-context calculation must use the profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/RedHatAI/gemma-4-26B-A4B-it-NVFP4" }, { "label": "Model config", "url": "https://huggingface.co/RedHatAI/gemma-4-26B-A4B-it-NVFP4/raw/24eef434bc1089b25f8866c687fd015c63546735/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/RedHatAI/gemma-4-26B-A4B-it-NVFP4" }, { "label": "Pinned Google base config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it/raw/20da991ab4afab98e8f910c4a2e8f4fbefc404ad/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/RedHatAI/gemma-4-26B-A4B-it-NVFP4/raw/24eef434bc1089b25f8866c687fd015c63546735/model.safetensors.index.json" } ] }, { "id": "deepseek-ai/DeepSeek-V2-Lite-Chat", "name": "DeepSeek-V2-Lite-Chat BF16", "short_name": "DeepSeek V2 Lite", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 1126300, "tags": [ "transformers", "safetensors", "deepseek_v2", "text-generation", "conversational", "custom_code", "arxiv:2405.04434", "license:other", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "deepseek_v2 (DeepseekV2ForCausalLM)", "total_params_b": 15.706484224, "active_params_b": 2.451435008, "layers": 27, "attention_heads": 16, "kv_heads": 16, "head_dim": 128, "max_context_tokens": 163840, "routed_experts": 64, "routed_experts_per_token": 6, "shared_experts_per_token": 2 }, "adapter": { "kind": "moe", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 31.412968448, "active_weight_gb": 4.902870016, "kv_alloc_gb_per_1k_tokens": 0.27648, "kv_read_gb_per_1k_tokens": 0.27648, "notes": "Manual correction on 2026-07-02 from current Hugging Face API metadata, model card, pinned config, pinned custom modeling code, safetensors index, direct range-read shard headers, and comparison with the audited DeepSeek-Coder-V2-Lite-Instruct profile. Bound weights separate the 31.412968448 GB resident BF16 checkpoint from 4.902870016 GB single-token active ordinary decode traffic; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite-Chat" }, { "label": "Model config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite-Chat/raw/85864749cd611b4353ce1decdb286193298f64c7/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-V2-Lite-Chat" }, { "label": "Pinned custom modeling code", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite-Chat/raw/85864749cd611b4353ce1decdb286193298f64c7/modeling_deepseek.py" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite-Chat/resolve/85864749cd611b4353ce1decdb286193298f64c7/model.safetensors.index.json" } ] }, { "id": "nm-testing/SmolLM-1.7B-Instruct-quantized.w4a16", "name": "SmolLM-1.7B-Instruct-quantized.w4a16 4-bit", "short_name": "SmolLM 1.7B Instruct quantized.w4a16", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 1061745, "tags": [ "safetensors", "llama", "text-generation", "conversational", "en", "arxiv:2210.17323", "license:apache-2.0", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "llama (LlamaForCausalLM)", "total_params_b": 1.83720584, "active_params_b": 1.736542544, "layers": 24, "attention_heads": 32, "kv_heads": 32, "head_dim": 64, "max_context_tokens": 2048, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "INT4/F32 compressed-tensors", "weight_bytes_per_param": 0.9316757495175391, "resident_weight_gb": 1.711680128, "active_weight_gb": 1.309026944, "kv_alloc_gb_per_1k_tokens": 0.196608, "kv_read_gb_per_1k_tokens": 0.196608, "notes": "Manually audited on 2026-07-02 from current Hugging Face API metadata, pinned config, model card, public SmolLM-1.7B-Instruct base config comparison, and direct safetensors header range read. Bound weights separate the 1.711680128 GB resident compressed-tensors package from 1.309026944 GB ordinary text-decode swept traffic; model.embed_tokens.weight is resident-only and lm_head.weight is swept. The audited profile supersedes the ideal flat 4-bit generated estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nm-testing/SmolLM-1.7B-Instruct-quantized.w4a16" }, { "label": "Model config", "url": "https://huggingface.co/nm-testing/SmolLM-1.7B-Instruct-quantized.w4a16/raw/1c917fc65b6cf47c20a2c2b4d6e2a8148fc20e9e/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nm-testing/SmolLM-1.7B-Instruct-quantized.w4a16" }, { "label": "Pinned base config", "url": "https://huggingface.co/HuggingFaceTB/SmolLM-1.7B-Instruct/raw/69f49d9c36434d6a3f319dadc5bd3b812752b98b/config.json" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/nm-testing/SmolLM-1.7B-Instruct-quantized.w4a16/resolve/1c917fc65b6cf47c20a2c2b4d6e2a8148fc20e9e/model.safetensors" } ] }, { "id": "deepseek-ai/DeepSeek-R1-0528-Qwen3-8B", "name": "DeepSeek-R1-0528-Qwen3-8B BF16", "short_name": "DeepSeek R1 0528 Qwen3 8B", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 1236831, "tags": [ "transformers", "safetensors", "qwen3", "text-generation", "conversational", "arxiv:2501.12948", "license:mit", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 8.19073536, "active_params_b": 7.568405504, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 16.38147072, "active_weight_gb": 15.136811008, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Manually audited on 2026-07-02 from current Hugging Face API metadata, pinned DeepSeek config, pinned model card, Qwen3-8B-Base config comparison, safetensors index, and direct range-read shard headers. Bound weights separate the 16.38147072 GB resident BF16 checkpoint from 15.136811008 GB ordinary text-decode swept traffic; model.embed_tokens.weight is resident-only and lm_head.weight is swept. The audited profile supersedes the generated flat BF16 estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-0528-Qwen3-8B" }, { "label": "Model config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-0528-Qwen3-8B/raw/6e8885a6ff5c1dc5201574c8fd700323f23c25fa/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-R1-0528-Qwen3-8B" }, { "label": "Pinned base config", "url": "https://huggingface.co/Qwen/Qwen3-8B-Base/raw/49e3418fbbbca6ecbdf9608b4d22e5a407081db4/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-0528-Qwen3-8B/raw/6e8885a6ff5c1dc5201574c8fd700323f23c25fa/model.safetensors.index.json" } ] }, { "id": "lmstudio-community/Qwen3.6-27B-MLX-4bit", "name": "Qwen3.6-27B-MLX-4bit 4-bit", "short_name": "Qwen3.6 27B MLX 4bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1132490, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "mlx", "conversational", "base_model:Qwen/Qwen3.6-27B", "base_model:quantized:Qwen/Qwen3.6-27B", "license:apache-2.0", "endpoints_compatible", "4-bit", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 27.35672856, "active_params_b": 25.624600064, "layers": 64, "attention_heads": 24, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "MLX 4-bit affine", "weight_bytes_per_param": 0.5868487602524941, "resident_weight_gb": 16.05426224, "active_weight_gb": 14.417640448, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Manually audited on 2026-07-02 from current Hugging Face API metadata, pinned served MLX config, Qwen3.6 27B base config comparison, model card, safetensors index, direct shard header range reads, and the Transformers qwen3_5 implementation. Bound weights separate the exact 16.05426224 GB resident MLX package from the 14.417640448 GB ordinary text-decode swept traffic; the audited profile supersedes the generated ideal 4-bit estimate and adds the fixed DeltaNet runtime state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-27B-MLX-4bit" }, { "label": "Model config", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-27B-MLX-4bit/raw/bd83f6fe15b171f1549475db2348389c0f541c21/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.6-27B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3.6-27B-MLX-4bit" }, { "label": "Pinned base config", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-27B-MLX-4bit/raw/bd83f6fe15b171f1549475db2348389c0f541c21/model.safetensors.index.json" } ] }, { "id": "cyankiwi/Qwen3.6-27B-AWQ-INT4", "name": "Qwen3.6-27B-AWQ-INT4 4-bit", "short_name": "Qwen3.6 27B AWQ INT4", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 953536, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "conversational", "base_model:Qwen/Qwen3.6-27B", "base_model:quantized:Qwen/Qwen3.6-27B", "license:apache-2.0", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 27.781427952, "active_params_b": 25.624600064, "layers": 64, "attention_heads": 24, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "compressed-tensors INT4", "weight_bytes_per_param": 0.7358756551794972, "resident_weight_gb": 20.443676496, "active_weight_gb": 16.659305728, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Manually audited on 2026-07-03 from current Hugging Face API metadata, pinned served compressed-tensors config, Qwen3.6 27B base config comparison, model card, safetensors index, direct shard header range reads, and the Transformers qwen3_5 implementation. Bound weights separate the exact 20.443676496 GB resident package from 16.659305728 GB ordinary text-decode swept traffic; the audited profile supersedes the generated ideal 4-bit estimate and adds the fixed DeltaNet runtime state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/cyankiwi/Qwen3.6-27B-AWQ-INT4" }, { "label": "Model config", "url": "https://huggingface.co/cyankiwi/Qwen3.6-27B-AWQ-INT4/raw/c9b937c5466c5c0575fc15edd1f8c516cb1e62fd/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.6-27B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3.6-27B-AWQ-INT4" }, { "label": "Pinned base config", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/cyankiwi/Qwen3.6-27B-AWQ-INT4/resolve/c9b937c5466c5c0575fc15edd1f8c516cb1e62fd/model.safetensors.index.json" } ] }, { "id": "lmstudio-community/Qwen3.6-27B-MLX-8bit", "name": "Qwen3.6-27B-MLX-8bit 8-bit", "short_name": "Qwen3.6 27B MLX 8bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1144822, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "mlx", "conversational", "base_model:Qwen/Qwen3.6-27B", "base_model:quantized:Qwen/Qwen3.6-27B", "license:apache-2.0", "endpoints_compatible", "8-bit", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 27.35672856, "active_params_b": 25.624600064, "layers": 64, "attention_heads": 24, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "MLX 8-bit affine", "weight_bytes_per_param": 1.0783796262516265, "resident_weight_gb": 29.50093872, "active_weight_gb": 27.228617728, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Manually audited on 2026-07-03 from current Hugging Face API metadata, pinned served MLX config, Qwen3.6 27B base config comparison, model card, safetensors index, direct shard header range reads, and the Transformers qwen3_5 implementation. Bound weights separate the exact 29.50093872 GB resident MLX package from the 27.228617728 GB ordinary text-decode swept traffic; the audited profile supersedes the generated ideal 8-bit estimate and adds the fixed DeltaNet runtime state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-27B-MLX-8bit" }, { "label": "Model config", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-27B-MLX-8bit/raw/75c2bb5358faad8cea31345d1112d79544a3d311/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.6-27B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3.6-27B-MLX-8bit" }, { "label": "Pinned base config", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-27B-MLX-8bit/raw/75c2bb5358faad8cea31345d1112d79544a3d311/model.safetensors.index.json" } ] }, { "id": "TheBloke/TinyLlama-1.1B-Chat-v0.3-GPTQ", "name": "TinyLlama-1.1B-Chat-v0.3-GPTQ 4-bit", "short_name": "TinyLlama 1.1B Chat v0.3", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 1068629, "tags": [ "transformers", "safetensors", "llama", "text-generation", "en", "dataset:cerebras/SlimPajama-627B", "dataset:bigcode/starcoderdata", "dataset:OpenAssistant/oasst_top1_2023-08-25", "base_model:TinyLlama/TinyLlama-1.1B-Chat-v0.3", "base_model:quantized:TinyLlama/TinyLlama-1.1B-Chat-v0.3", "license:apache-2.0", "text-generation-inference", "4-bit", "gptq", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "llama (LlamaForCausalLM)", "total_params_b": 1.100454912, "active_params_b": 1.034912768, "layers": 22, "attention_heads": 32, "kv_heads": 4, "head_dim": 64, "max_context_tokens": 2048, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GPTQ 4-bit group 128 act-order", "weight_bytes_per_param": 0.6979695029976839, "resident_weight_gb": 0.768083968, "active_weight_gb": 0.63699968, "kv_alloc_gb_per_1k_tokens": 0.022528, "kv_read_gb_per_1k_tokens": 0.022528, "notes": "Manually audited on 2026-07-03 from current Hugging Face API metadata, model card, pinned GPTQ config, quantize_config, current TinyLlama base config comparison, and direct safetensors header range reads. Bound weights separate the exact 0.768083968 GB resident GPTQ package from the 0.63699968 GB ordinary text-decode swept traffic; the audited profile supersedes the generated ideal 4-bit estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v0.3-GPTQ" }, { "label": "Model config", "url": "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v0.3-GPTQ/raw/05835c52707fff57cecd16a364de2bc65c9bf102/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v0.3" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/TheBloke/TinyLlama-1.1B-Chat-v0.3-GPTQ" }, { "label": "Quantize config", "url": "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v0.3-GPTQ/raw/05835c52707fff57cecd16a364de2bc65c9bf102/quantize_config.json" }, { "label": "Pinned base config", "url": "https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v0.3/raw/1ea93498c83d6d5aece90cf3b160aadf9f8875f8/config.json" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v0.3-GPTQ/resolve/05835c52707fff57cecd16a364de2bc65c9bf102/model.safetensors" } ] }, { "id": "Qwen/Qwen2.5-72B-Instruct-AWQ", "name": "Qwen2.5-72B-Instruct-AWQ 4-bit", "short_name": "Qwen2.5 72B Instruct", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 1056283, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "chat", "conversational", "en", "arxiv:2309.00071", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-72B-Instruct", "base_model:quantized:Qwen/Qwen2.5-72B-Instruct", "license:other", "text-generation-inference", "endpoints_compatible", "4-bit", "awq", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 72.957861888, "active_params_b": 71.7121536, "layers": 80, "attention_heads": 64, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "AWQ 4-bit GEMM", "weight_bytes_per_param": 0.570130728883676, "resident_weight_gb": 41.595518976, "active_weight_gb": 39.1041024, "kv_alloc_gb_per_1k_tokens": 0.32768, "kv_read_gb_per_1k_tokens": 0.32768, "notes": "Manually audited on 2026-07-03 from current Hugging Face API metadata, model card, pinned served AWQ config, current Qwen2.5 72B Instruct base config comparison, safetensors index, and direct shard header range reads. Bound weights separate the exact 41.595518976 GB resident AWQ package from the 39.1041024 GB ordinary text-decode swept traffic; the audited profile supersedes the generated ideal 4-bit estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-72B-Instruct-AWQ" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-72B-Instruct-AWQ/raw/698703eae6604af048a3d2f509995dc302088217/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-72B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-72B-Instruct-AWQ" }, { "label": "Pinned base config", "url": "https://huggingface.co/Qwen/Qwen2.5-72B-Instruct/raw/495f39366efef23836d0cfae4fbe635880d2be31/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Qwen/Qwen2.5-72B-Instruct-AWQ/raw/698703eae6604af048a3d2f509995dc302088217/model.safetensors.index.json" } ] }, { "id": "unsloth/Qwen3.5-9B-GGUF", "name": "Qwen3.5-9B-GGUF GGUF", "short_name": "Qwen3.5 9B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1040131, "tags": [ "transformers", "gguf", "unsloth", "image-text-to-text", "base_model:Qwen/Qwen3.5-9B", "base_model:quantized:Qwen/Qwen3.5-9B", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text (selected BF16 GGUF)", "total_params_b": 8.953803264, "active_params_b": 8.953803264, "layers": 32, "attention_heads": 16, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16/F32 GGUF", "weight_bytes_per_param": 2.0002370782490315, "resident_weight_gb": 17.920697312, "active_weight_gb": 15.87549184, "kv_alloc_gb_per_1k_tokens": 0.032768, "kv_read_gb_per_1k_tokens": 0.032768, "kv_alloc_fixed_gb": 0.051904512, "kv_read_fixed_gb": 0.051904512, "notes": "Manually audited on 2026-07-03 from HF API commit 3885219b6810b007914f3a7950a8d1b469d598a5, Qwen base config, model card, linked GGUF file sizes, and a GGUF tensor-index range read. The authoritative bounds profile uses the selected Qwen3.5-9B-BF16.gguf artifact: 17.920697312 GB resident, 15.87549184 GB swept text traffic, 2.045205472 GB resident-only token embedding plus GGUF metadata/header/file overhead, and Qwen3.5 hybrid KV plus fixed DeltaNet state. Separate mmproj GGUF files are not included unless explicitly loaded." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/Qwen3.5-9B-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3.5-9B/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-9B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3.5-9B-GGUF" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/unsloth/Qwen3.5-9B-GGUF/resolve/3885219b6810b007914f3a7950a8d1b469d598a5/Qwen3.5-9B-BF16.gguf" } ] }, { "id": "Qwen/Qwen2-0.5B", "name": "Qwen2-0.5B BF16", "short_name": "Qwen2 0.5B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 1017191, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "pretrained", "conversational", "en", "license:apache-2.0", "eval-results", "text-generation-inference", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 0.494032768, "active_params_b": 0.494032768, "layers": 24, "attention_heads": 14, "kv_heads": 2, "head_dim": 64, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 0.988065536, "active_weight_gb": 0.988065536, "kv_alloc_gb_per_1k_tokens": 0.012288, "kv_read_gb_per_1k_tokens": 0.012288, "notes": "Manually audited on 2026-07-03 from HF API commit 91d2aff3f957f99e4c74c962f2f408dcc88a18d8, pinned config, model card, and a safetensors header range read. The authoritative bounds profile records 494032768 BF16 parameters / 0.988065536 GB, tied embeddings with no separate lm_head.weight, and full-context Qwen2 KV over 24 layers." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2-0.5B" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2-0.5B/raw/91d2aff3f957f99e4c74c962f2f408dcc88a18d8/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2-0.5B" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/Qwen/Qwen2-0.5B/resolve/91d2aff3f957f99e4c74c962f2f408dcc88a18d8/model.safetensors" } ] }, { "id": "cyankiwi/gemma-4-31B-it-AWQ-4bit", "name": "gemma-4-31B-it-AWQ-4bit 4-bit", "short_name": "gemma 4 31B it AWQ 4bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1035024, "tags": [ "transformers", "safetensors", "image-text-to-text", "base_model:google/gemma-4-31B-it", "base_model:finetune:google/gemma-4-31B-it", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text (Gemma4ForConditionalGeneration, AWQ)", "total_params_b": 33.103510176, "active_params_b": 33.103510176, "layers": 60, "attention_heads": 32, "kv_heads": 16, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "AWQ 4-bit", "weight_bytes_per_param": 0.6314773956254179, "resident_weight_gb": 20.904118392, "active_weight_gb": 19.75263132, "kv_alloc_gb_per_1k_tokens": 0.83968, "kv_read_gb_per_1k_tokens": 0.83968, "notes": "Manually audited on 2026-07-03 from HF API commit 853ac120bc4d1cd4aa7ae66a412f76a35c4aca4f, pinned AWQ config, Google base config comparison, and range-read safetensors shard headers. The authoritative bounds profile records exact resident/swept stored bytes, resident-only vision/embed_vision bytes, no separate lm_head.weight, and layered Gemma 4 KV; the legacy KV fields here are short-context unsaturated approximations only." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/cyankiwi/gemma-4-31B-it-AWQ-4bit" }, { "label": "Model config", "url": "https://huggingface.co/cyankiwi/gemma-4-31B-it-AWQ-4bit/raw/853ac120bc4d1cd4aa7ae66a412f76a35c4aca4f/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-31B-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/cyankiwi/gemma-4-31B-it-AWQ-4bit" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/cyankiwi/gemma-4-31B-it-AWQ-4bit/raw/853ac120bc4d1cd4aa7ae66a412f76a35c4aca4f/model.safetensors.index.json" } ] }, { "id": "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4", "name": "NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4 NVFP4", "short_name": "NVIDIA Nemotron 3 Super 120B A12B NVFP4", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 946756, "tags": [ "transformers", "safetensors", "nemotron_h", "text-generation", "nvidia", "pytorch", "nemotron-3", "latent-moe", "mtp", "conversational", "custom_code", "en", "fr", "es", "it", "de", "ja", "zh", "dataset:nvidia/nemotron-post-training-v3", "dataset:nvidia/nemotron-pre-training-datasets", "arxiv:2512.20848", "arxiv:2512.20856", "license:other", "endpoints_compatible", "8-bit", "modelopt", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "nemotron_h (NemotronHForCausalLM)", "total_params_b": 120, "active_params_b": 12, "layers": 88, "attention_heads": 32, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 512, "routed_experts_per_token": 22, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "NVFP4/FP8/BF16", "weight_bytes_per_param": 0.5, "resident_weight_gb": 80.297329824, "active_weight_gb": 12.64573328, "kv_alloc_gb_per_1k_tokens": 0.004096, "kv_read_gb_per_1k_tokens": 0.004096, "kv_alloc_fixed_gb": 0.08716288, "kv_read_fixed_gb": 0.08716288, "expert_param_b": 0.24772736, "fixed_param_b": 19.84146464, "shared_expert_param_b": 0, "notes": "Manual correction on 2026-07-03 from current Hugging Face API metadata, pinned served config, BF16 config comparison, ModelOpt quantization config, model card serving notes, custom runtime review, safetensors index, and direct range-read safetensors shard headers. Legacy fallback values mirror the audited profile: exact stored resident bytes, fixed always-on hybrid/shared-expert traffic, expected distinct routed experts, FP8 attention KV, and fixed float16-sized Mamba recurrent state. The authoritative profile supersedes the generated 33.6143 GB resident estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4" }, { "label": "Pinned served config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4/raw/4f0cf9daaeb7a4d5e23f80a00e7ed15f0e03caf6/config.json" }, { "label": "Pinned ModelOpt quantization config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4/raw/4f0cf9daaeb7a4d5e23f80a00e7ed15f0e03caf6/hf_quant_config.json" }, { "label": "Pinned BF16 config comparison", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16/raw/d51eab0d1f979ebc26b546e634a04f450d99158e/config.json" }, { "label": "Custom runtime code", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4/raw/4f0cf9daaeb7a4d5e23f80a00e7ed15f0e03caf6/modeling_nemotron_h.py" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4/resolve/4f0cf9daaeb7a4d5e23f80a00e7ed15f0e03caf6/model.safetensors.index.json" } ] }, { "id": "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16", "name": "NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 F32/BF16", "short_name": "NVIDIA Nemotron 3 Nano 30B A3B BF16", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 1064387, "tags": [ "transformers", "safetensors", "nemotron_h", "text-generation", "nvidia", "pytorch", "conversational", "custom_code", "en", "es", "fr", "de", "ja", "it", "dataset:nvidia/Nemotron-Pretraining-Code-v1", "dataset:nvidia/Nemotron-CC-v2", "dataset:nvidia/Nemotron-Pretraining-SFT-v1", "dataset:nvidia/Nemotron-CC-Math-v1", "dataset:nvidia/Nemotron-Pretraining-Code-v2", "dataset:nvidia/Nemotron-Pretraining-Specialized-v1", "dataset:nvidia/Nemotron-CC-v2.1", "dataset:nvidia/Nemotron-CC-Code-v1", "dataset:nvidia/Nemotron-Pretraining-Dataset-sample", "dataset:nvidia/Nemotron-Competitive-Programming-v1", "dataset:nvidia/Nemotron-Math-v2", "dataset:nvidia/Nemotron-Agentic-v1", "dataset:nvidia/Nemotron-Math-Proofs-v1", "dataset:nvidia/Nemotron-Instruction-Following-Chat-v1", "dataset:nvidia/Nemotron-Science-v1", "dataset:nvidia/Nemotron-3-Nano-RL-Training-Blend", "arxiv:2512.20848", "arxiv:2512.20856", "license:other", "eval-results", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "nemotron_h (NemotronHForCausalLM)", "total_params_b": 31.577937344, "active_params_b": 3.227760704, "layers": 52, "attention_heads": 32, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 6, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "F32/BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 63.155892352, "active_weight_gb": 6.455521408, "kv_alloc_gb_per_1k_tokens": 0.006144, "kv_read_gb_per_1k_tokens": 0.006144, "kv_alloc_fixed_gb": 0.049364992, "kv_read_fixed_gb": 0.049364992, "expert_param_b": 0.229490688, "fixed_param_b": 1.850816576, "shared_expert_param_b": 0, "notes": "Manual correction on 2026-07-03 from current Hugging Face API metadata, pinned served config, base config comparison, custom runtime review, safetensors index, and direct range-read safetensors shard headers. Legacy fallback values mirror the audited profile: exact stored resident bytes, fixed always-on hybrid/shared-expert traffic, expected distinct routed experts, BF16 attention KV, and fixed BF16-convolution/F32-SSM Mamba recurrent state. This supersedes the generated all-layer KV estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16" }, { "label": "Pinned served config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/raw/cbd3fa9f933d55ef16a84236559f4ee2a0526848/config.json" }, { "label": "Pinned base config comparison", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-Base-BF16/raw/97ab8012882a655dc38df4fee47422aca9caca07/config.json" }, { "label": "Custom runtime code", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/raw/cbd3fa9f933d55ef16a84236559f4ee2a0526848/modeling_nemotron_h.py" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/resolve/cbd3fa9f933d55ef16a84236559f4ee2a0526848/model.safetensors.index.json" } ] }, { "id": "RedHatAI/Qwen2.5-1.5B-quantized.w8a8", "name": "Qwen2.5-1.5B-quantized.w8a8 8-bit", "short_name": "Qwen2.5 1.5B quantized.w8a8", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 1009174, "tags": [ "safetensors", "qwen2", "chat", "neuralmagic", "llmcompressor", "text-generation", "conversational", "en", "base_model:Qwen/Qwen2.5-1.5B", "base_model:quantized:Qwen/Qwen2.5-1.5B", "license:apache-2.0", "8-bit", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 1.77773312, "active_params_b": 1.544359424, "layers": 28, "attention_heads": 12, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "INT8/BF16", "weight_bytes_per_param": 1, "resident_weight_gb": 2.245270528, "active_weight_gb": 1.778523136, "kv_alloc_gb_per_1k_tokens": 0.028672, "kv_read_gb_per_1k_tokens": 0.028672, "notes": "Manually audited on 2026-07-03 from current HF API metadata, pinned RedHatAI config, model card, recipe, Qwen source-snapshot config comparison, and a direct safetensors header range read. Bound weights separate the 2.245270528 GB resident compressed-tensors package from the 1.778523136 GB ordinary text-decode swept subset; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/RedHatAI/Qwen2.5-1.5B-quantized.w8a8" }, { "label": "Pinned model config", "url": "https://huggingface.co/RedHatAI/Qwen2.5-1.5B-quantized.w8a8/raw/45d7501a9947504178b3d1bfee36272bd1582e66/config.json" }, { "label": "Pinned source-snapshot config", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B/raw/cf341214a7c12d36db6c7fbbf8b113c7ed61502f/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/RedHatAI/Qwen2.5-1.5B-quantized.w8a8" }, { "label": "Quantization recipe", "url": "https://huggingface.co/RedHatAI/Qwen2.5-1.5B-quantized.w8a8/raw/45d7501a9947504178b3d1bfee36272bd1582e66/recipe.yaml" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/RedHatAI/Qwen2.5-1.5B-quantized.w8a8/resolve/45d7501a9947504178b3d1bfee36272bd1582e66/model.safetensors" } ] }, { "id": "unsloth/Qwen3.5-4B-GGUF", "name": "Qwen3.5-4B-GGUF BF16 GGUF", "short_name": "Qwen3.5 4B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1058880, "tags": [ "transformers", "gguf", "unsloth", "image-text-to-text", "base_model:Qwen/Qwen3.5-4B", "base_model:quantized:Qwen/Qwen3.5-4B", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text (selected BF16 GGUF)", "total_params_b": 4.205751296, "active_params_b": 4.205751296, "layers": 32, "attention_heads": 16, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16/F32 GGUF", "weight_bytes_per_param": 2.0004572481501293, "resident_weight_gb": 8.424393632, "active_weight_gb": 8.413425664, "kv_alloc_gb_per_1k_tokens": 0.032768, "kv_read_gb_per_1k_tokens": 0.032768, "kv_alloc_fixed_gb": 0.051904512, "kv_read_fixed_gb": 0.051904512, "notes": "Manually audited on 2026-07-03 from HF API commit e87f176479d0855a907a41277aca2f8ee7a09523, Qwen base config, model card, linked GGUF file sizes, and a GGUF tensor-index range read. The authoritative bounds profile uses the selected Qwen3.5-4B-BF16.gguf artifact: 8.424393632 GB resident, 8.413425664 GB swept text traffic, 0.010967968 GB resident-only GGUF metadata/header/file overhead, and Qwen3.5 hybrid KV plus fixed DeltaNet state. The tied token embedding is swept because no separate output.weight is stored. Separate mmproj GGUF files are not included unless explicitly loaded." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/Qwen3.5-4B-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3.5-4B/raw/851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-4B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3.5-4B-GGUF" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/unsloth/Qwen3.5-4B-GGUF/resolve/e87f176479d0855a907a41277aca2f8ee7a09523/Qwen3.5-4B-BF16.gguf" } ] }, { "id": "google/gemma-3-27b-it", "name": "gemma-3-27b-it BF16", "short_name": "gemma 3 27b it", "license": "gemma", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 962269, "tags": [ "transformers", "safetensors", "gemma3", "image-text-to-text", "conversational", "arxiv:1905.07830", "arxiv:1905.10044", "arxiv:1911.11641", "arxiv:1904.09728", "arxiv:1705.03551", "arxiv:1911.01547", "arxiv:1907.10641", "arxiv:1903.00161", "arxiv:2009.03300", "arxiv:2304.06364", "arxiv:2103.03874", "arxiv:2110.14168", "arxiv:2311.12022", "arxiv:2108.07732", "arxiv:2107.03374", "arxiv:2210.03057", "arxiv:2106.03193", "arxiv:1910.11856", "arxiv:2502.12404", "arxiv:2502.21228", "arxiv:2404.16816", "arxiv:2104.12756", "arxiv:2311.16502", "arxiv:2203.10244", "arxiv:2404.12390", "arxiv:1810.12440", "arxiv:1908.02660", "arxiv:2312.11805", "base_model:google/gemma-3-27b-pt", "base_model:finetune:google/gemma-3-27b-pt", "license:gemma", "eval-results", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense multimodal transformer", "detail": "Gemma3ForConditionalGeneration (gated raw config)", "total_params_b": 27.43240664, "active_params_b": 27.43240664, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 54.86481328, "active_weight_gb": 54.86481328, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manually corrected on 2026-07-03 from current public Hugging Face API metadata. The repo is gated; raw config, README, generation config, safetensors index, base-repo config/index, ranged safetensors bytes, and hf download config.json all returned access denied. Production bounds use the unsupported profile and must not infer KV geometry from this row." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-3-27b-it" }, { "label": "Base model config", "url": "https://huggingface.co/google/gemma-3-27b-pt/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-3-27b-pt" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-3-27b-it" } ] }, { "id": "lmstudio-community/Qwen3.6-27B-MLX-6bit", "name": "Qwen3.6-27B-MLX-6bit 6-bit", "short_name": "Qwen3.6 27B MLX 6bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1070949, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "mlx", "conversational", "base_model:Qwen/Qwen3.6-27B", "base_model:quantized:Qwen/Qwen3.6-27B", "license:apache-2.0", "endpoints_compatible", "6-bit", "region:us" ], "architecture": { "type": "Dense multimodal transformer", "detail": "qwen3_5_text", "total_params_b": 27.35672856, "active_params_b": 25.624600064, "layers": 64, "attention_heads": 24, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "MLX 6-bit affine", "weight_bytes_per_param": 0.8326141932520604, "resident_weight_gb": 22.77760048, "active_weight_gb": 20.823129088, "kv_alloc_gb_per_1k_tokens": 0.065536, "kv_read_gb_per_1k_tokens": 0.065536, "kv_alloc_fixed_gb": 0.154927104, "kv_read_fixed_gb": 0.154927104, "notes": "Manually audited on 2026-07-03 from current Hugging Face API metadata, pinned served MLX config, Qwen3.6 27B base config comparison, model card, safetensors index, direct shard header range reads, and the Transformers qwen3_5 implementation. Bound weights separate the exact 22.77760048 GB resident MLX package from the 20.823129088 GB ordinary text-decode swept traffic; the audited profile supersedes the generated BF16/U32 dense estimate and adds the fixed DeltaNet runtime state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-27B-MLX-6bit" }, { "label": "Model config", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-27B-MLX-6bit/raw/dafc96bab96a89e3ea8c4b3c197e2806b8c0d595/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.6-27B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3.6-27B-MLX-6bit" }, { "label": "Pinned base config", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-27B-MLX-6bit/raw/dafc96bab96a89e3ea8c4b3c197e2806b8c0d595/model.safetensors.index.json" } ] }, { "id": "lmstudio-community/Qwen3.6-27B-MLX-5bit", "name": "Qwen3.6-27B-MLX-5bit 5-bit", "short_name": "Qwen3.6 27B MLX 5bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1065162, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "mlx", "conversational", "base_model:Qwen/Qwen3.6-27B", "base_model:quantized:Qwen/Qwen3.6-27B", "license:apache-2.0", "endpoints_compatible", "5-bit", "region:us" ], "architecture": { "type": "Dense multimodal transformer", "detail": "qwen3_5_text", "total_params_b": 27.35672856, "active_params_b": 25.624600064, "layers": 64, "attention_heads": 24, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "MLX 5-bit affine", "weight_bytes_per_param": 0.7097314767522773, "resident_weight_gb": 19.41593136, "active_weight_gb": 17.620384768, "kv_alloc_gb_per_1k_tokens": 0.065536, "kv_read_gb_per_1k_tokens": 0.065536, "kv_alloc_fixed_gb": 0.154927104, "kv_read_fixed_gb": 0.154927104, "notes": "Manually audited on 2026-07-03 from current Hugging Face API metadata, pinned served MLX config, Qwen3.6 27B base config comparison, model card, safetensors index, direct shard header range reads, and the Transformers qwen3_5 implementation. Bound weights separate the exact 19.41593136 GB resident MLX package from the 17.620384768 GB ordinary text-decode swept traffic; the audited profile supersedes the generated BF16/U32 dense estimate and adds the fixed DeltaNet runtime state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-27B-MLX-5bit" }, { "label": "Model config", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-27B-MLX-5bit/raw/589f789ab2a6673feaf18e2acaa0867b07f657d8/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.6-27B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3.6-27B-MLX-5bit" }, { "label": "Pinned base config", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-27B-MLX-5bit/raw/589f789ab2a6673feaf18e2acaa0867b07f657d8/model.safetensors.index.json" } ] }, { "id": "empero-ai/Qwythos-9B-Claude-Mythos-5-1M-GGUF", "name": "Qwythos-9B-Claude-Mythos-5-1M-GGUF Q4_K_M", "short_name": "Qwythos 9B Claude Mythos 5 1M", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1250562, "tags": [ "gguf", "llama.cpp", "quantized", "qwen3.5", "reasoning", "uncensored", "long-context", "1M-context", "function-calling", "multimodal", "vision", "cybersecurity", "biomedical", "agentic", "image-text-to-text", "en", "base_model:empero-ai/Qwythos-9B-Claude-Mythos-5-1M", "base_model:quantized:empero-ai/Qwythos-9B-Claude-Mythos-5-1M", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense multimodal transformer", "detail": "qwen3_5_text (card-recommended Q4_K_M GGUF)", "total_params_b": 8.953803264, "active_params_b": 8.953803264, "layers": 32, "attention_heads": 16, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 1048576, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q4_K_M", "weight_bytes_per_param": 0.6286835868543827, "resident_weight_gb": 5.629109152, "active_weight_gb": 5.046011904, "kv_alloc_gb_per_1k_tokens": 0.032768, "kv_read_gb_per_1k_tokens": 0.032768, "kv_alloc_fixed_gb": 0.051904512, "kv_read_fixed_gb": 0.051904512, "notes": "Manually audited on 2026-07-03 from HF API commit a7ab0f6ee807c165e8374e4906773ca39f5fdff3, model card, base config, linked GGUF file sizes, and a GGUF tensor-index range read. The authoritative bounds profile uses the card-recommended normal text Q4_K_M artifact: 5.629109152 GB resident, 5.046011904 GB swept text traffic, 0.583097248 GB resident-only token embedding plus GGUF metadata/header/file overhead, 1048576-token YaRN context, and Qwen3.5 hybrid KV plus fixed DeltaNet state. Separate MTP and mmproj GGUF sidecars are not included unless explicitly loaded." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/empero-ai/Qwythos-9B-Claude-Mythos-5-1M-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/empero-ai/Qwythos-9B-Claude-Mythos-5-1M/raw/763f72fc2c3b186e977adcbaac0c18128f182166/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/empero-ai/Qwythos-9B-Claude-Mythos-5-1M" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/empero-ai/Qwythos-9B-Claude-Mythos-5-1M-GGUF" }, { "label": "Selected GGUF artifact", "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" } ] }, { "id": "zai-org/GLM-5.1-FP8", "name": "GLM-5.1-FP8 FP8", "short_name": "GLM 5.1 FP8", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 938761, "tags": [ "transformers", "safetensors", "glm_moe_dsa", "text-generation", "conversational", "en", "zh", "arxiv:2602.15763", "license:mit", "eval-results", "endpoints_compatible", "fp8", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "glm_moe_dsa (GlmMoeDsaForCausalLM)", "total_params_b": 753.910024032, "active_params_b": 23.5597, "layers": 78, "attention_heads": 64, "kv_heads": 64, "head_dim": 256, "max_context_tokens": 202752, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "FP8", "weight_bytes_per_param": 1.0029879743844663, "resident_weight_gb": 756.162687872, "active_weight_gb": 41.928982272, "kv_alloc_gb_per_1k_tokens": 5.131776, "kv_read_gb_per_1k_tokens": 5.131776, "notes": "Manually corrected on 2026-07-03 from current Hugging Face API metadata, pinned config comparison with GLM-5-FP8, direct safetensors header range reads across all 142 shards, and the Transformers GLM MoE DSA cache path. Active traffic is the single-session expected-distinct MoE text-decode weight traffic; KV display fields charge expanded BF16 K/V plus the DSA indexer key cache. The audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/zai-org/GLM-5.1-FP8" }, { "label": "Model config", "url": "https://huggingface.co/zai-org/GLM-5.1-FP8/raw/f396cf805182f4ca10fa675e1a99815b3ca384db/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/zai-org/GLM-5.1-FP8" }, { "label": "Safetensors index", "url": "https://huggingface.co/zai-org/GLM-5.1-FP8/resolve/f396cf805182f4ca10fa675e1a99815b3ca384db/model.safetensors.index.json" } ] }, { "id": "nvidia/DeepSeek-R1-0528-NVFP4-v2", "name": "DeepSeek-R1-0528-NVFP4-v2 NVFP4", "short_name": "DeepSeek R1 0528 NVFP4 v2", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 930601, "tags": [ "Model Optimizer", "safetensors", "deepseek_v3", "nvidia", "ModelOpt", "DeepSeekR1", "quantized", "FP4", "text-generation", "conversational", "custom_code", "base_model:deepseek-ai/DeepSeek-R1-0528", "base_model:finetune:deepseek-ai/DeepSeek-R1-0528", "license:mit", "8-bit", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "deepseek_v3 (DeepseekV3ForCausalLM)", "total_params_b": 684.5314, "active_params_b": 21.3916, "layers": 61, "attention_heads": 128, "kv_heads": 128, "head_dim": 56, "max_context_tokens": 163840, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "NVFP4 with FP8 scales and BF16 fixed tensors", "weight_bytes_per_param": 1, "resident_weight_gb": 413.306995616, "active_weight_gb": 28.197245728, "kv_alloc_gb_per_1k_tokens": 2.49856, "kv_read_gb_per_1k_tokens": 2.49856, "expert_param_b": 1.436812656, "shared_expert_param_b": 0, "fixed_param_b": 16.70274448, "notes": "Audited on 2026-07-03 at commit 25a138f28f49022958b9f2d205f9b7de0cdb6e18; downloads=930601. Exact profile data supersedes the earlier metadata estimate: stored tensors are range-read from safetensors headers, model.embed_tokens.weight plus layer 61 are resident-only for ordinary decode, and FP8 expanded K/V cache traffic is charged from the shipped runtime path." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/DeepSeek-R1-0528-NVFP4-v2" }, { "label": "Model config", "url": "https://huggingface.co/nvidia/DeepSeek-R1-0528-NVFP4-v2/raw/25a138f28f49022958b9f2d205f9b7de0cdb6e18/config.json" }, { "label": "ModelOpt quantization config", "url": "https://huggingface.co/nvidia/DeepSeek-R1-0528-NVFP4-v2/raw/25a138f28f49022958b9f2d205f9b7de0cdb6e18/hf_quant_config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/DeepSeek-R1-0528-NVFP4-v2" }, { "label": "Safetensors index", "url": "https://huggingface.co/nvidia/DeepSeek-R1-0528-NVFP4-v2/resolve/25a138f28f49022958b9f2d205f9b7de0cdb6e18/model.safetensors.index.json" } ] }, { "id": "deepseek-ai/deepseek-vl2-tiny", "name": "deepseek-vl2-tiny BF16", "short_name": "deepseek vl2 tiny", "license": "other", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 985805, "tags": [ "transformers", "safetensors", "deepseek_vl_v2", "image-text-to-text", "arxiv:2412.10302", "license:other", "endpoints_compatible", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "deepseek_vl_v2 with DeepseekV2 language model", "total_params_b": 3.37050144, "active_params_b": 0.57412736, "layers": 12, "attention_heads": 10, "kv_heads": 10, "head_dim": 128, "max_context_tokens": 4096, "routed_experts": 64, "routed_experts_per_token": 6, "shared_experts_per_token": 2 }, "adapter": { "kind": "moe", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 6.74100288, "active_weight_gb": 1.14825472, "kv_alloc_gb_per_1k_tokens": 0.06144, "kv_read_gb_per_1k_tokens": 0.06144, "expert_param_b": 0.03784704, "shared_expert_param_b": 0, "fixed_param_b": 0.34704512, "notes": "Manually audited on 2026-07-03 from current Hugging Face API metadata, pinned config, model card, direct safetensors header range read, and pinned DeepSeek-VL2 runtime files. Bound weights use exact BF16 tensor bytes, separate resident-only vision/projector/image/input-embedding tensors from ordinary text decode, and charge standard BF16 full-context MHA KV because the language config has use_mla false." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/deepseek-ai/deepseek-vl2-tiny" }, { "label": "Model config", "url": "https://huggingface.co/deepseek-ai/deepseek-vl2-tiny/raw/66c54660eae7e90c9ba259bfdf92d07d6e3ce8aa/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/deepseek-vl2-tiny" }, { "label": "Safetensors index", "url": "https://huggingface.co/deepseek-ai/deepseek-vl2-tiny/resolve/66c54660eae7e90c9ba259bfdf92d07d6e3ce8aa/model.safetensors.index.json" }, { "label": "DeepSeek-VL2 runtime", "url": "https://raw.githubusercontent.com/deepseek-ai/DeepSeek-VL2/ef9f91e2b6426536b83294c11742c27be66361b1/deepseek_vl2/models/modeling_deepseek_vl_v2.py" } ] }, { "id": "rippertnt/HyperCLOVAX-SEED-Text-Instruct-1.5B-Q4_K_M-GGUF", "name": "HyperCLOVAX-SEED-Text-Instruct-1.5B-Q4_K_M-GGUF GGUF 4-bit", "short_name": "HyperCLOVAX SEED Text Instruct 1.5B Q4 K M", "license": "other", "hf_pipeline_tag": "unknown", "hf_downloads": 913990, "tags": [ "gguf", "llama", "llama-cpp", "gguf-my-repo", "base_model:naver-hyperclovax/HyperCLOVAX-SEED-Text-Instruct-1.5B", "base_model:quantized:naver-hyperclovax/HyperCLOVAX-SEED-Text-Instruct-1.5B", "license:other", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "llama (LlamaForCausalLM, Q4_K_M GGUF)", "total_params_b": 1.81203968, "active_params_b": 1.81203968, "layers": 24, "attention_heads": 16, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q4_K_M", "weight_bytes_per_param": 0.6257999648219624, "resident_weight_gb": 1.133974368, "active_weight_gb": 1.002512384, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "notes": "Manually audited on 2026-07-03 from HF API commit 3d2edd543d75372922aabd4ae538e6d181634b5d, model card, quantized repo config, base-model API metadata, linked GGUF object checks, and a GGUF tensor-index range read. The authoritative bounds profile uses the selected Q4_K_M artifact: 1.133974368 GB resident, 1.002512384 GB swept text traffic, 0.131461984 GB resident-only token embedding plus GGUF metadata/header/file overhead, 131072-token context, and full-context FP16 K/V cache. The base repo raw config was gated during audit." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/rippertnt/HyperCLOVAX-SEED-Text-Instruct-1.5B-Q4_K_M-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/rippertnt/HyperCLOVAX-SEED-Text-Instruct-1.5B-Q4_K_M-GGUF/raw/3d2edd543d75372922aabd4ae538e6d181634b5d/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/rippertnt/HyperCLOVAX-SEED-Text-Instruct-1.5B-Q4_K_M-GGUF" }, { "label": "Base model API metadata", "url": "https://huggingface.co/api/models/naver-hyperclovax/HyperCLOVAX-SEED-Text-Instruct-1.5B" }, { "label": "Selected GGUF artifact", "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" } ] }, { "id": "Qwen/Qwen3.5-35B-A3B-GPTQ-Int4", "name": "Qwen3.5-35B-A3B-GPTQ-Int4 4-bit", "short_name": "Qwen3.5 35B A3B GPTQ Int4", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 915845, "tags": [ "transformers", "safetensors", "qwen3_5_moe", "image-text-to-text", "conversational", "base_model:Qwen/Qwen3.5-35B-A3B", "base_model:quantized:Qwen/Qwen3.5-35B-A3B", "license:apache-2.0", "endpoints_compatible", "4-bit", "gptq", "deploy:azure", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_5_moe_text GPTQ Int4", "total_params_b": 35.951822704, "active_params_b": 3, "layers": 40, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "GPTQ Int4 packed I32 + BF16/F16 side tensors", "weight_bytes_per_param": 0.6787739917643982, "resident_weight_gb": 24.403162208, "active_weight_gb": 4.408478336, "kv_alloc_gb_per_1k_tokens": 0.02048, "kv_read_gb_per_1k_tokens": 0.02048, "kv_alloc_fixed_gb": 0.06488064, "kv_read_fixed_gb": 0.06488064, "expert_param_b": 0.09739536399760368, "fixed_param_b": 5.715603224447949, "shared_expert_param_b": 0, "notes": "Manually audited on 2026-07-03 from HF API commit 3af5ca2972faf6de1fd6f4efc4d8d319ca751e8b, config.json, base config comparison, model card, and range-read safetensors headers. The authoritative bounds profile uses exact stored bytes: 24.403162208 GB resident, 3.879602816 GB fixed text traffic, 0.06610944 GB per routed expert, 8 routed experts plus 1 shared expert, 10 full-attention layers, and a fixed DeltaNet recurrent-state charge of 0.06488064 GB per session/token." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B-GPTQ-Int4" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B-GPTQ-Int4/raw/3af5ca2972faf6de1fd6f4efc4d8d319ca751e8b/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-35B-A3B-GPTQ-Int4" }, { "label": "Safetensors index", "url": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B-GPTQ-Int4/resolve/3af5ca2972faf6de1fd6f4efc4d8d319ca751e8b/model.safetensors.index.json" } ] }, { "id": "zai-org/GLM-5.2-FP8", "name": "GLM-5.2-FP8 FP8", "short_name": "GLM 5.2 FP8", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 1051277, "tags": [ "transformers", "safetensors", "glm_moe_dsa", "text-generation", "conversational", "en", "zh", "arxiv:2602.15763", "arxiv:2603.12201", "license:mit", "endpoints_compatible", "fp8", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "glm_moe_dsa (GlmMoeDsaForCausalLM)", "total_params_b": 753.375793584, "active_params_b": 23.543, "layers": 78, "attention_heads": 64, "kv_heads": 64, "head_dim": 256, "max_context_tokens": 1048576, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "FP8", "weight_bytes_per_param": 1.002975071473079, "resident_weight_gb": 755.617140416, "active_weight_gb": 41.383434816, "kv_alloc_gb_per_1k_tokens": 5.117184, "kv_read_gb_per_1k_tokens": 5.117184, "expert_param_b": 2.823446445025638, "fixed_param_b": 18.6731097797805, "shared_expert_param_b": 0, "notes": "Manually corrected on 2026-07-03 from current Hugging Face API metadata, pinned config comparison with GLM-5-FP8 and GLM-5.1-FP8, direct safetensors header range reads across all 141 shards, and the Transformers GLM MoE DSA config/runtime/cache path. Active traffic is the single-session expected-distinct MoE text-decode weight traffic. KV display fields charge expanded BF16 K/V for 78 layers plus GLM-5.2 IndexShare DSA indexer key cache for 21 full indexer layers. The audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/zai-org/GLM-5.2-FP8" }, { "label": "Model config", "url": "https://huggingface.co/zai-org/GLM-5.2-FP8/raw/ba978f7d347eaf65d22f1a86833408afdb953541/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/zai-org/GLM-5.2-FP8" }, { "label": "Safetensors index", "url": "https://huggingface.co/zai-org/GLM-5.2-FP8/resolve/ba978f7d347eaf65d22f1a86833408afdb953541/model.safetensors.index.json" } ] }, { "id": "unsloth/Qwen3.6-27B-MTP-GGUF", "name": "Qwen3.6-27B-MTP-GGUF GGUF", "short_name": "Qwen3.6 27B MTP", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 946357, "tags": [ "transformers", "gguf", "unsloth", "qwen", "qwen3_5", "image-text-to-text", "base_model:Qwen/Qwen3.6-27B", "base_model:quantized:Qwen/Qwen3.6-27B", "license:apache-2.0", "endpoints_compatible", "region:us", "imatrix", "conversational" ], "architecture": { "type": "Dense multimodal transformer", "detail": "qwen3_5_text (card-example UD-Q4_K_XL GGUF)", "total_params_b": 27.320697856, "active_params_b": 25.624600064, "layers": 64, "attention_heads": 24, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF UD-Q4_K_XL", "weight_bytes_per_param": 0.6555139145564292, "resident_weight_gb": 17.9090976, "active_weight_gb": 16.916391936, "kv_alloc_gb_per_1k_tokens": 0.065536, "kv_read_gb_per_1k_tokens": 0.065536, "kv_alloc_fixed_gb": 0.154927104, "kv_read_fixed_gb": 0.154927104, "notes": "Manually audited on 2026-07-03 from HF API commit 5cb35eb3dcbf52dbce5f87dbc64df6aaffadcace, model card, pinned Qwen base config, linked GGUF file sizes, and a GGUF tensor-index range read. The authoritative bounds profile uses the card-example Qwen3.6-27B-UD-Q4_K_XL.gguf artifact: 17.9090976 GB resident, 16.916391936 GB swept ordinary text traffic, 0.992705664 GB resident-only token embedding plus blk.64 MTP plus GGUF metadata/header/file overhead, and Qwen3.6 hybrid KV plus fixed DeltaNet state. The HF API gguf.totalFileSize points to IQ4_NL, so the selected artifact is intentionally pinned by the model card's llama.cpp MTP command." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/Qwen3.6-27B-MTP-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.6-27B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3.6-27B-MTP-GGUF" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/unsloth/Qwen3.6-27B-MTP-GGUF/resolve/5cb35eb3dcbf52dbce5f87dbc64df6aaffadcace/Qwen3.6-27B-UD-Q4_K_XL.gguf" } ] }, { "id": "RedHatAI/Llama-3.2-1B-Instruct-FP8", "name": "Llama-3.2-1B-Instruct-FP8 FP8", "short_name": "Llama 3.2 1B Instruct FP8", "license": "llama3.2", "hf_pipeline_tag": "text-generation", "hf_downloads": 884137, "tags": [ "safetensors", "llama", "llama-3", "neuralmagic", "llmcompressor", "text-generation", "conversational", "en", "de", "fr", "it", "pt", "hi", "es", "th", "base_model:meta-llama/Llama-3.2-1B-Instruct", "base_model:quantized:meta-llama/Llama-3.2-1B-Instruct", "license:llama3.2", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "llama (LlamaForCausalLM)", "total_params_b": 1.498482912, "active_params_b": 1.235814624, "layers": 16, "attention_heads": 32, "kv_heads": 8, "head_dim": 64, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "FP8", "weight_bytes_per_param": 1, "resident_weight_gb": 2.023887296, "active_weight_gb": 1.49855072, "kv_alloc_gb_per_1k_tokens": 0.032768, "kv_read_gb_per_1k_tokens": 0.032768, "notes": "Manually audited on 2026-07-03 from current Hugging Face API metadata, model card, served static FP8 config, recipe.yaml, and direct safetensors header range read. Bound weights separate the 2.023887296 GB resident compressed-tensors package from the 1.49855072 GB ordinary text-decode swept subset; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/RedHatAI/Llama-3.2-1B-Instruct-FP8" }, { "label": "Model config", "url": "https://huggingface.co/RedHatAI/Llama-3.2-1B-Instruct-FP8/raw/fb49430ca7e61099fa1ff30f12fecb290a8ebb65/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/RedHatAI/Llama-3.2-1B-Instruct-FP8" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/RedHatAI/Llama-3.2-1B-Instruct-FP8/resolve/fb49430ca7e61099fa1ff30f12fecb290a8ebb65/model.safetensors" } ] }, { "id": "Qwen/Qwen3-4B-Instruct-2507-FP8", "name": "Qwen3-4B-Instruct-2507-FP8 FP8", "short_name": "Qwen3 4B Instruct 2507 FP8", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 888237, "tags": [ "transformers", "safetensors", "qwen3", "text-generation", "conversational", "arxiv:2505.09388", "base_model:Qwen/Qwen3-4B-Instruct-2507", "base_model:quantized:Qwen/Qwen3-4B-Instruct-2507", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "fp8", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 4.411646016, "active_params_b": 4.022689856, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "FP8", "weight_bytes_per_param": 1, "resident_weight_gb": 5.189976192, "active_weight_gb": 4.412063872, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Manually audited on 2026-07-03 from current Hugging Face API metadata, model card, served FP8 config, base profile comparison, and direct safetensors header range read. Bound weights separate the 5.189976192 GB resident FP8/F16 package from the 4.412063872 GB ordinary text-decode swept subset; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507-FP8" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507-FP8/raw/8591804019c8b22094c3b5b4454e0edc05dffc98/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-4B-Instruct-2507-FP8" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507-FP8/resolve/8591804019c8b22094c3b5b4454e0edc05dffc98/model.safetensors" } ] }, { "id": "Qwen/Qwen2.5-1.5B-Instruct-AWQ", "name": "Qwen2.5-1.5B-Instruct-AWQ 4-bit", "short_name": "Qwen2.5 1.5B Instruct", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 863664, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "chat", "conversational", "en", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-1.5B-Instruct", "base_model:quantized:Qwen/Qwen2.5-1.5B-Instruct", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "4-bit", "awq", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 1.777088, "active_params_b": 1.543714304, "layers": 28, "attention_heads": 12, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "AWQ 4-bit GEMM", "weight_bytes_per_param": 0.9084931033240897, "resident_weight_gb": 1.614472192, "active_weight_gb": 1.1477248, "kv_alloc_gb_per_1k_tokens": 0.028672, "kv_read_gb_per_1k_tokens": 0.028672, "notes": "Manually audited on 2026-07-03 from current Hugging Face API metadata, model card, pinned served AWQ config, base Instruct config comparison, and direct safetensors header range read. Bound weights separate the exact 1.614472192 GB resident AWQ package from the 1.147724800 GB ordinary text-decode swept traffic; the audited profile supersedes the generated ideal 4-bit estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct-AWQ" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct-AWQ/raw/3ecffa0ceb27851800f45519bab9c457a04405e1/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-1.5B-Instruct-AWQ" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct-AWQ/resolve/3ecffa0ceb27851800f45519bab9c457a04405e1/model.safetensors" } ] }, { "id": "unsloth/Qwen3.6-35B-A3B-GGUF", "name": "Qwen3.6-35B-A3B-GGUF GGUF", "short_name": "Qwen3.6 35B A3B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 868905, "tags": [ "transformers", "gguf", "unsloth", "qwen", "qwen3_5_moe", "image-text-to-text", "base_model:Qwen/Qwen3.6-35B-A3B", "base_model:quantized:Qwen/Qwen3.6-35B-A3B", "license:apache-2.0", "endpoints_compatible", "deploy:azure", "region:us", "imatrix", "conversational" ], "architecture": { "type": "MoE transformer", "detail": "qwen35moe / qwen3_5_moe_text", "total_params_b": 34.660610688, "active_params_b": 3, "layers": 40, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "GGUF MXFP4_MOE", "weight_bytes_per_param": 0.626248190817797, "resident_weight_gb": 21.706144736, "active_weight_gb": 2.73211648, "kv_alloc_gb_per_1k_tokens": 0.02048, "kv_read_gb_per_1k_tokens": 0.02048, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, Qwen base config, selected linked-object HEAD check, direct GGUF header range read, and ggml MXFP4 type definitions. The selected API artifact is Qwen3.6-35B-A3B-MXFP4_MOE.gguf at 21.706144736 GB resident; ordinary single-session swept traffic is 2.732116480 GB before KV/state reads. The audited profile adds the fixed DeltaNet state charge that these legacy per-1K KV fields cannot fully express." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/Qwen3.6-35B-A3B-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3.6-35B-A3B-GGUF" }, { "label": "Selected GGUF header audit", "url": "https://huggingface.co/unsloth/Qwen3.6-35B-A3B-GGUF/resolve/a483e9e6cbd595906af30beda3187c2663a1118c/Qwen3.6-35B-A3B-MXFP4_MOE.gguf" }, { "label": "GGUF MXFP4 type definitions", "url": "https://github.com/ggml-org/ggml/blob/master/src/ggml-common.h" } ] }, { "id": "nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16", "name": "Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16 F32/BF16", "short_name": "Nemotron 3 Nano Omni 30B A3B Reasoning BF16", "license": "other", "hf_pipeline_tag": "any-to-any", "hf_downloads": 1003984, "tags": [ "transformers", "safetensors", "NemotronH_Nano_Omni_Reasoning_V3", "feature-extraction", "nvidia", "pytorch", "multimodal", "any-to-any", "custom_code", "dataset:nvidia/Nemotron-Image-Training-v3", "arxiv:2604.24954", "license:other", "deploy:azure", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "NemotronH_Nano_Omni_Reasoning_V3 / nemotron_h", "total_params_b": 33.015632214, "active_params_b": 3.22775776, "layers": 52, "attention_heads": 32, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 6, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "F32/BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 66.03127052, "active_weight_gb": 6.45551552, "kv_alloc_gb_per_1k_tokens": 0.006144, "kv_read_gb_per_1k_tokens": 0.006144, "kv_alloc_fixed_gb": 0.024870912, "kv_read_fixed_gb": 0.024870912, "expert_param_b": 0.229490688, "fixed_param_b": 1.850813632, "shared_expert_param_b": 0, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, model card, pinned BF16 Omni config, standalone BF16 backbone config comparison, custom runtime review, safetensors index, and direct range-read safetensors shard headers. Legacy fallback values mirror the audited profile: exact stored resident bytes, fixed always-on hybrid/shared-expert traffic, expected distinct routed experts, BF16 attention KV, and fixed BF16 Mamba recurrent state. This supersedes the generated all-layer KV estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16" }, { "label": "Pinned BF16 Omni config", "url": "https://huggingface.co/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16/raw/24e67ea000b7c2837fc8f9488aa2008524fac8ba/config.json" }, { "label": "Standalone BF16 backbone config comparison", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/raw/cbd3fa9f933d55ef16a84236559f4ee2a0526848/config.json" }, { "label": "Custom runtime code", "url": "https://huggingface.co/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16/raw/24e67ea000b7c2837fc8f9488aa2008524fac8ba/modeling_nemotron_h.py" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16/resolve/24e67ea000b7c2837fc8f9488aa2008524fac8ba/model.safetensors.index.json" } ] }, { "id": "unsloth/Qwen3-VL-2B-Instruct-GGUF", "name": "Qwen3-VL-2B-Instruct-GGUF GGUF", "short_name": "Qwen3 VL 2B Instruct", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 856615, "tags": [ "transformers", "gguf", "unsloth", "image-text-to-text", "arxiv:2505.09388", "arxiv:2502.13923", "arxiv:2409.12191", "arxiv:2308.12966", "base_model:Qwen/Qwen3-VL-2B-Instruct", "base_model:quantized:Qwen/Qwen3-VL-2B-Instruct", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen3vl / qwen3_vl_text (selected BF16 GGUF)", "total_params_b": 1.720574976, "active_params_b": 1.720574976, "layers": 28, "attention_heads": 16, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16/F32 GGUF", "weight_bytes_per_param": 2.0001440262722965, "resident_weight_gb": 3.44735072, "active_weight_gb": 3.44139776, "kv_alloc_gb_per_1k_tokens": 0.114688, "kv_read_gb_per_1k_tokens": 0.114688, "notes": "Manually audited on 2026-07-05 from HF API commit 8dcb98e52a1d1d02dce9249e5ab15bae8121c666, Qwen base config, model card, linked GGUF file sizes, and a GGUF tensor-index range read. The authoritative bounds profile uses the selected Qwen3-VL-2B-Instruct-BF16.gguf artifact: 3.447350720 GB resident, 3.441397760 GB swept text traffic, 0.005952960 GB resident-only GGUF metadata/header/file overhead, and full-context Qwen3-VL text KV. The tied token embedding is swept because no separate output.weight is stored. Separate mmproj GGUF files are not included unless explicitly loaded." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/Qwen3-VL-2B-Instruct-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3-VL-2B-Instruct/raw/89644892e4d85e24eaac8bacfd4f463576704203/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-2B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3-VL-2B-Instruct-GGUF" }, { "label": "Linked-object HEAD checks", "url": "https://huggingface.co/unsloth/Qwen3-VL-2B-Instruct-GGUF/tree/8dcb98e52a1d1d02dce9249e5ab15bae8121c666" }, { "label": "Selected BF16 GGUF header audit", "url": "https://huggingface.co/unsloth/Qwen3-VL-2B-Instruct-GGUF/resolve/8dcb98e52a1d1d02dce9249e5ab15bae8121c666/Qwen3-VL-2B-Instruct-BF16.gguf" } ] }, { "id": "nvidia/NVIDIA-Nemotron-Nano-9B-v2", "name": "NVIDIA-Nemotron-Nano-9B-v2 BF16", "short_name": "NVIDIA Nemotron Nano 9B v2", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 728827, "tags": [ "transformers", "safetensors", "nemotron_h", "text-generation", "nvidia", "pytorch", "conversational", "custom_code", "en", "es", "fr", "de", "it", "ja", "dataset:nvidia/Nemotron-Post-Training-Dataset-v1", "dataset:nvidia/Nemotron-Post-Training-Dataset-v2", "dataset:nvidia/Nemotron-Pretraining-Dataset-sample", "dataset:nvidia/Nemotron-CC-v2", "dataset:nvidia/Nemotron-CC-Math-v1", "dataset:nvidia/Nemotron-Pretraining-SFT-v1", "arxiv:2504.03624", "arxiv:2508.14444", "arxiv:2412.02595", "base_model:nvidia/NVIDIA-Nemotron-Nano-12B-v2", "base_model:finetune:nvidia/NVIDIA-Nemotron-Nano-12B-v2", "license:other", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "Hybrid Mamba/Transformer", "detail": "nemotron_h (NemotronHForCausalLM); 27 Mamba, 4 attention, 25 MLP layers", "total_params_b": 8.888227328, "active_params_b": 8.301024768, "layers": 56, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "hybrid", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 17.776454656, "active_weight_gb": 16.602049536, "kv_alloc_gb_per_1k_tokens": 0.016384, "kv_read_gb_per_1k_tokens": 0.016384, "kv_alloc_fixed_gb": 0.144211968, "kv_read_fixed_gb": 0.144211968, "notes": "Manually corrected on 2026-07-05 from current Hugging Face API metadata, model card, served config, 12B parent config comparison, custom runtime code, safetensors index, and direct range-read safetensors shard headers. The repo currently records BF16:8888227328, a 56-layer hybrid Mamba2/Transformer pattern with 27 Mamba layers, 4 full-attention layers, and 25 MLP layers. vLLM serving requires F32 Mamba SSM cache for accuracy, so legacy fallback values include fixed BF16-convolution/F32-SSM state. The audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2" }, { "label": "Model config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2/raw/6533e8de2c68e4536bf7c411d7a3ce5734111476/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-12B-v2-Base" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/NVIDIA-Nemotron-Nano-9B-v2" }, { "label": "Custom runtime code", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2/raw/6533e8de2c68e4536bf7c411d7a3ce5734111476/modeling_nemotron_h.py" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2/resolve/6533e8de2c68e4536bf7c411d7a3ce5734111476/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen3.5-397B-A17B-FP8", "name": "Qwen3.5-397B-A17B-FP8 FP8", "short_name": "Qwen3.5 397B A17B FP8", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 829107, "tags": [ "transformers", "safetensors", "qwen3_5_moe", "image-text-to-text", "conversational", "base_model:Qwen/Qwen3.5-397B-A17B", "base_model:quantized:Qwen/Qwen3.5-397B-A17B", "license:apache-2.0", "endpoints_compatible", "fp8", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_5_moe_text", "total_params_b": 403.422386928, "active_params_b": 17, "layers": 60, "attention_heads": 32, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "FP8 F8_E4M3 + BF16 side tensors", "weight_bytes_per_param": 1, "resident_weight_gb": 406.12518128, "active_weight_gb": 17.503303552, "kv_alloc_gb_per_1k_tokens": 0.03072, "kv_read_gb_per_1k_tokens": 0.03072, "kv_alloc_fixed_gb": 0.19316736, "kv_read_fixed_gb": 0.19316736, "notes": "Manually audited on 2026-07-05 from HF API commit ea5b4f81096f3901c91dea97f81324302495781d, config.json, base config comparison, and range-read safetensors headers. The authoritative bounds profile uses exact stored bytes: 406.12518128 GB resident, 9.952634752 GB fixed text traffic, 0.75506688 GB per routed expert, 10 routed experts plus 1 shared expert, 15 full-attention layers, and a fixed DeltaNet recurrent-state charge of 0.19316736 GB per session/token." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3.5-397B-A17B-FP8" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3.5-397B-A17B-FP8/raw/ea5b4f81096f3901c91dea97f81324302495781d/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-397B-A17B" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3.5-397B-A17B/raw/8472618112abcbd45acbcdc58436aff4233c23f7/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-397B-A17B-FP8" }, { "label": "Safetensors index", "url": "https://huggingface.co/Qwen/Qwen3.5-397B-A17B-FP8/resolve/ea5b4f81096f3901c91dea97f81324302495781d/model.safetensors.index.json" } ] }, { "id": "microsoft/Phi-tiny-MoE-instruct", "name": "Phi-tiny-MoE-instruct BF16", "short_name": "Phi tiny MoE", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 884992, "tags": [ "transformers", "safetensors", "phimoe", "text-generation", "conversational", "custom_code", "en", "arxiv:2506.18349", "arxiv:2404.14219", "arxiv:2409.12136", "license:mit", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "phimoe (PhiMoEForCausalLM)", "total_params_b": 3.755220288, "active_params_b": 1.1, "layers": 32, "attention_heads": 16, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 4096, "routed_experts": 16, "routed_experts_per_token": 2, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 7.510440576, "active_weight_gb": 2.315270784, "kv_alloc_gb_per_1k_tokens": 0.065536, "kv_read_gb_per_1k_tokens": 0.065536, "notes": "Manually audited on 2026-07-05 from HF API commit 2fe50e88d0e2a5a132563815686ea0dcc8e252b5, model card, config.json, remote SlimMoE modeling code, and range-read safetensors headers. The authoritative bounds profile uses exact stored bytes: 7.510440576 GB resident, 1.610627712 GB fixed text traffic, 0.352321536 GB per routed expert, 2 routed experts, and BF16 K/V traffic capped by the 2047-token sliding window." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/microsoft/Phi-tiny-MoE-instruct" }, { "label": "Model config", "url": "https://huggingface.co/microsoft/Phi-tiny-MoE-instruct/raw/2fe50e88d0e2a5a132563815686ea0dcc8e252b5/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/microsoft/Phi-tiny-MoE-instruct" }, { "label": "Custom SlimMoE runtime code", "url": "https://huggingface.co/microsoft/Phi-tiny-MoE-instruct/raw/2fe50e88d0e2a5a132563815686ea0dcc8e252b5/modeling_slimmoe.py" }, { "label": "Safetensors index", "url": "https://huggingface.co/microsoft/Phi-tiny-MoE-instruct/resolve/2fe50e88d0e2a5a132563815686ea0dcc8e252b5/model.safetensors.index.json" } ] }, { "id": "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4", "name": "NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4 NVFP4", "short_name": "NVIDIA Nemotron 3 Nano 30B A3B NVFP4", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 1094976, "tags": [ "transformers", "safetensors", "nemotron_h", "text-generation", "nvidia", "pytorch", "conversational", "custom_code", "en", "es", "fr", "de", "ja", "it", "dataset:nvidia/Nemotron-Pretraining-Code-v1", "dataset:nvidia/Nemotron-CC-v2", "dataset:nvidia/Nemotron-Pretraining-SFT-v1", "dataset:nvidia/Nemotron-CC-Math-v1", "dataset:nvidia/Nemotron-Pretraining-Code-v2", "dataset:nvidia/Nemotron-Pretraining-Specialized-v1", "dataset:nvidia/Nemotron-CC-v2.1", "dataset:nvidia/Nemotron-CC-Code-v1", "dataset:nvidia/Nemotron-Pretraining-Dataset-sample", "dataset:nvidia/Nemotron-Competitive-Programming-v1", "dataset:nvidia/Nemotron-Math-v2", "dataset:nvidia/Nemotron-Agentic-v1", "dataset:nvidia/Nemotron-Math-Proofs-v1", "dataset:nvidia/Nemotron-Instruction-Following-Chat-v1", "dataset:nvidia/Nemotron-Science-v1", "dataset:nvidia/Nemotron-3-Nano-RL-Training-Blend", "arxiv:2512.20848", "arxiv:2512.20856", "arxiv:2601.20088", "base_model:nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16", "base_model:quantized:nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16", "license:other", "endpoints_compatible", "8-bit", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "nemotron_h (NemotronHForCausalLM)", "total_params_b": 30, "active_params_b": 3.5, "layers": 52, "attention_heads": 32, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 6, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "NVFP4/FP8/BF16", "weight_bytes_per_param": 0.5, "resident_weight_gb": 19.339781632, "active_weight_gb": 2.8862952, "kv_alloc_gb_per_1k_tokens": 0.003072, "kv_read_gb_per_1k_tokens": 0.003072, "kv_alloc_fixed_gb": 0.049364992, "kv_read_fixed_gb": 0.049364992, "expert_param_b": 0.25817776, "fixed_param_b": 4.22352384, "shared_expert_param_b": 0, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, pinned served config, BF16 config comparison, ModelOpt quantization config, model card serving notes, custom runtime review, safetensors index, and direct range-read safetensors shard headers. Legacy fallback values mirror the audited profile: exact stored resident bytes, fixed always-on hybrid/shared-expert traffic, expected distinct routed experts, FP8 attention KV, and fixed BF16-convolution/F32-SSM Mamba recurrent state. The authoritative profile supersedes the generated 15.7889 GB resident estimate and all-layer KV estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4" }, { "label": "Pinned served config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4/raw/ce1b118ae66ec705d02c241525192832eb045fd3/config.json" }, { "label": "Pinned ModelOpt quantization config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4/raw/ce1b118ae66ec705d02c241525192832eb045fd3/hf_quant_config.json" }, { "label": "Pinned BF16 config comparison", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/raw/cbd3fa9f933d55ef16a84236559f4ee2a0526848/config.json" }, { "label": "Custom runtime code", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4/raw/ce1b118ae66ec705d02c241525192832eb045fd3/modeling_nemotron_h.py" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4/resolve/ce1b118ae66ec705d02c241525192832eb045fd3/model.safetensors.index.json" } ] }, { "id": "microsoft/phi-4", "name": "phi-4 BF16", "short_name": "phi 4", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 869780, "tags": [ "transformers", "safetensors", "phi3", "text-generation", "phi", "nlp", "math", "code", "chat", "conversational", "en", "arxiv:2412.08905", "license:mit", "eval-results", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "phi3 (Phi3ForCausalLM)", "total_params_b": 14.6595072, "active_params_b": 14.14570496, "layers": 40, "attention_heads": 40, "kv_heads": 10, "head_dim": 128, "max_context_tokens": 16384, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 29.3190144, "active_weight_gb": 28.29140992, "kv_alloc_gb_per_1k_tokens": 0.2048, "kv_read_gb_per_1k_tokens": 0.2048, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, model card, pinned config, official Transformers 4.47.0 Phi3 implementation, safetensors index, and direct range-read shard headers. Bound weights separate the 29.319014400 GB resident BF16 checkpoint from the 28.291409920 GB ordinary text-decode swept subset; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/microsoft/phi-4" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/microsoft/phi-4" }, { "label": "Pinned model config", "url": "https://huggingface.co/microsoft/phi-4/raw/932b33c0ec9ca189badeb22480721a8de9d0e006/config.json" }, { "label": "Transformers Phi3 implementation", "url": "https://raw.githubusercontent.com/huggingface/transformers/v4.47.0/src/transformers/models/phi3/modeling_phi3.py" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/microsoft/phi-4/resolve/932b33c0ec9ca189badeb22480721a8de9d0e006/model.safetensors.index.json" }, { "label": "License", "url": "https://huggingface.co/microsoft/phi-4/raw/932b33c0ec9ca189badeb22480721a8de9d0e006/LICENSE" } ] }, { "id": "unsloth/gemma-4-26B-A4B-it-qat-GGUF", "name": "gemma-4-26B-A4B-it-qat-GGUF GGUF 4-bit", "short_name": "gemma 4 26B A4B it qat", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 942587, "tags": [ "transformers", "gguf", "gemma4", "image-text-to-text", "unsloth", "gemma", "google", "base_model:google/gemma-4-26B-A4B-it-qat-q4_0-unquantized", "base_model:quantized:google/gemma-4-26B-A4B-it-qat-q4_0-unquantized", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "MoE transformer", "detail": "gemma4_text", "total_params_b": 25.233142046, "active_params_b": 3.8, "layers": 30, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "GGUF Q4_0 QAT", "weight_bytes_per_param": 0.5646956329902951, "resident_weight_gb": 14.24904512, "active_weight_gb": 2.189754304, "kv_alloc_gb_per_1k_tokens": 0.20992, "kv_read_gb_per_1k_tokens": 0.20992, "expert_param_b": 0.1777280257482084, "fixed_param_b": 2.4559360883597177, "shared_expert_param_b": 0, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, pinned model card, packaged config, base config comparison, linked GGUF file sizes, and a direct GGUF tensor-index range read. The authoritative bounds profile uses the selected gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf artifact: 14.249045120 GB resident, 1.386856384 GB fixed traffic, 0.100362240 GB per routed expert, 8 routed experts per token, 128 routed experts, 1 shared expert charged in fixed traffic, and layered Gemma 4 hybrid KV. Legacy per-1k KV fields show the short-context unsaturated coefficient; long-context calculation must use the profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/gemma-4-26B-A4B-it-qat-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/gemma-4-26B-A4B-it-qat-GGUF" }, { "label": "Pinned model config", "url": "https://huggingface.co/unsloth/gemma-4-26B-A4B-it-qat-GGUF/raw/02749a7b272109255a4c559a80894d3d9777574c/config.json" }, { "label": "Pinned base config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it-qat-q4_0-unquantized/raw/641f184470aa8554ae7957599a624badc2bf4e57/config.json" }, { "label": "GGUF linked-object and header audit", "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" } ] }, { "id": "google/gemma-4-31B-it-qat-w4a16-ct", "name": "gemma-4-31B-it-qat-w4a16-ct 4-bit", "short_name": "gemma 4 31B it qat w4a16 ct", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 987445, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "conversational", "base_model:google/gemma-4-31B-it-qat-q4_0-unquantized", "base_model:quantized:google/gemma-4-31B-it-qat-q4_0-unquantized", "license:apache-2.0", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 33.59758608, "active_params_b": 31.6125564, "layers": 60, "attention_heads": 32, "kv_heads": 16, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "compressed-tensors W4A16 int4", "weight_bytes_per_param": 0.6924630092353349, "resident_weight_gb": 23.26508556, "active_weight_gb": 19.2950262, "kv_alloc_gb_per_1k_tokens": 0.83968, "kv_read_gb_per_1k_tokens": 0.83968, "notes": "Manual correction on 2026-07-05 from live Hugging Face API metadata, served compressed-tensors config, unquantized QAT base config comparison, linked safetensors object metadata, and a direct safetensors header range read. The authoritative bounds profile uses exact stored bytes: 23.26508556 GB resident, 19.2950262 GB swept ordinary text traffic, 3.97005936 GB resident-only embedding and vision traffic, and layered Gemma 4 BF16 KV with separate sliding-layer K/V streams. Legacy per-1k KV fields show the short-context unsaturated coefficient; long-context calculation must use the profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-4-31B-it-qat-w4a16-ct" }, { "label": "Model config", "url": "https://huggingface.co/google/gemma-4-31B-it-qat-w4a16-ct/raw/ca9d5250d08695be35f2d632b434938b6d4aecfe/config.json" }, { "label": "Pinned base config", "url": "https://huggingface.co/google/gemma-4-31B-it-qat-q4_0-unquantized/raw/4f926903562062220b3e54c1385c5ef2cd40bfd1/config.json" }, { "label": "Safetensors linked-object and header audit", "url": "https://huggingface.co/google/gemma-4-31B-it-qat-w4a16-ct/resolve/ca9d5250d08695be35f2d632b434938b6d4aecfe/model.safetensors" } ] }, { "id": "Qwen/Qwen3-235B-A22B", "name": "Qwen3-235B-A22B BF16", "short_name": "Qwen3 235B A22B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 916836, "tags": [ "transformers", "safetensors", "qwen3_moe", "text-generation", "conversational", "arxiv:2309.00071", "arxiv:2505.09388", "license:apache-2.0", "eval-results", "endpoints_compatible", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_moe (Qwen3MoeForCausalLM)", "total_params_b": 235.09363456, "active_params_b": 22, "layers": 94, "attention_heads": 64, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 470.18726912, "active_weight_gb": 43.136867328, "kv_alloc_gb_per_1k_tokens": 0.192512, "kv_read_gb_per_1k_tokens": 0.192512, "notes": "Manual correction on 2026-07-05 from live Hugging Face API metadata, served config, model card, local scrape row, and direct safetensors header range reads for all 118 shards. The authoritative bounds profile uses exact stored bytes: 470.187269120 GB resident, 14.749817856 GB fixed ordinary text traffic, 3.548381184 GB per routed expert, 8 routed experts per token, 128 routed experts, no shared expert, and full-context BF16 KV." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-235B-A22B" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-235B-A22B/raw/8efa61729e24bd65b1d152b5ab5409052aa80e65/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-235B-A22B" }, { "label": "Pinned safetensors index and header audit", "url": "https://huggingface.co/Qwen/Qwen3-235B-A22B/raw/8efa61729e24bd65b1d152b5ab5409052aa80e65/model.safetensors.index.json" } ] }, { "id": "nvidia/LocateAnything-3B", "name": "LocateAnything-3B BF16", "short_name": "LocateAnything 3B", "license": "other", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 1194542, "tags": [ "transformers", "safetensors", "locateanything", "feature-extraction", "nvidia", "eagle", "vision", "object-detection", "grounding", "arxiv:2605.27365", "image-text-to-text", "conversational", "custom_code", "en", "arxiv:2504.07491", "arxiv:2109.10852", "arxiv:2510.12798", "arxiv:2303.05499", "arxiv:1405.0312", "arxiv:1908.03195", "arxiv:2504.07981", "base_model:Qwen/Qwen2.5-3B-Instruct", "base_model:finetune:Qwen/Qwen2.5-3B-Instruct", "license:other", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "locateanything custom Qwen2 grounding decoder", "total_params_b": 3.830665968, "active_params_b": 3.400155136, "layers": 36, "attention_heads": 16, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 7.661331936, "active_weight_gb": 6.800310272, "kv_alloc_gb_per_1k_tokens": 0.036864, "kv_read_gb_per_1k_tokens": 0.036864, "notes": "Manual correction on 2026-07-05 from live Hugging Face API metadata, served config, custom modeling/generation code, model card, and direct safetensors header range reads for both shards. The package has 7.661331936 GB BF16 resident tensors, including 6.800310272 GB language-model tensors and 0.861021664 GB vision/projector/input-embedding auxiliary tensors. Production bounds are disabled in the model profile because LocateAnything uses custom hybrid MTP/AR parallel box decoding and range-attention behavior outside Bounds Engine v1." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/LocateAnything-3B" }, { "label": "Model config", "url": "https://huggingface.co/nvidia/LocateAnything-3B/raw/c32291ca5e996f5a7a485845b4f57a233936bba0/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/LocateAnything-3B" }, { "label": "Pinned safetensors index and header audit", "url": "https://huggingface.co/nvidia/LocateAnything-3B/raw/c32291ca5e996f5a7a485845b4f57a233936bba0/model.safetensors.index.json" } ] }, { "id": "cyankiwi/Qwen3.5-9B-AWQ-4bit", "name": "Qwen3.5-9B-AWQ-4bit 4-bit", "short_name": "Qwen3.5 9B AWQ 4bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 758044, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "conversational", "base_model:Qwen/Qwen3.5-9B", "base_model:quantized:Qwen/Qwen3.5-9B", "license:apache-2.0", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 9.653104368, "active_params_b": 7.936684544, "layers": 32, "attention_heads": 16, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "compressed-tensors AWQ INT4", "weight_bytes_per_param": 0.9394325313690631, "resident_weight_gb": 9.068440272, "active_weight_gb": 5.937066112, "kv_alloc_gb_per_1k_tokens": 0.032768, "kv_read_gb_per_1k_tokens": 0.032768, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, pinned served compressed-tensors config, Qwen3.5 9B base config comparison, recipe, safetensors index, direct shard header range reads, and the Transformers qwen3_5 implementation. Bound weights separate the exact 9.068440272 GB resident package from 5.937066112 GB ordinary text-decode swept traffic; the audited profile supersedes the generated ideal 4-bit estimate and adds the fixed DeltaNet runtime state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/cyankiwi/Qwen3.5-9B-AWQ-4bit" }, { "label": "Model config", "url": "https://huggingface.co/cyankiwi/Qwen3.5-9B-AWQ-4bit/raw/73536aa464f9a93c550aa5a916f0113a08b2f384/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-9B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3.5-9B-AWQ-4bit" }, { "label": "Pinned base config", "url": "https://huggingface.co/Qwen/Qwen3.5-9B/raw/c202236235762e1c871ad0ccb60c8ee5ba337b9a/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/cyankiwi/Qwen3.5-9B-AWQ-4bit/resolve/73536aa464f9a93c550aa5a916f0113a08b2f384/model.safetensors.index.json" } ] }, { "id": "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B", "name": "DeepSeek-R1-Distill-Qwen-32B BF16", "short_name": "DeepSeek R1 Distill Qwen 32B", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 826902, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "conversational", "arxiv:2501.12948", "license:mit", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 32.763876352, "active_params_b": 31.985308672, "layers": 64, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 65.527752704, "active_weight_gb": 63.970617344, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, pinned model card, pinned served config, Qwen2.5 32B base config comparison, safetensors index, and direct shard header range reads. Bound weights separate the exact 65.527752704 GB resident package from 63.970617344 GB ordinary text-decode swept traffic; the audited profile supersedes the generated flat dense estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B" }, { "label": "Model config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B/raw/711ad2ea6aa40cfca18895e8aca02ab92df1a746/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B" }, { "label": "Pinned base config", "url": "https://huggingface.co/Qwen/Qwen2.5-32B/raw/1818d35814b8319459f4bd55ed1ac8709630f003/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B/raw/711ad2ea6aa40cfca18895e8aca02ab92df1a746/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen2.5-Coder-1.5B-Instruct", "name": "Qwen2.5-Coder-1.5B-Instruct BF16", "short_name": "Qwen2.5 Coder 1.5B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 805428, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "code", "codeqwen", "chat", "qwen", "qwen-coder", "conversational", "en", "arxiv:2409.12186", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-Coder-1.5B", "base_model:finetune:Qwen/Qwen2.5-Coder-1.5B", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 1.5437, "active_params_b": 1.5437, "layers": 28, "attention_heads": 12, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 3.087428608, "active_weight_gb": 3.087428608, "kv_alloc_gb_per_1k_tokens": 0.028672, "kv_read_gb_per_1k_tokens": 0.028672, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, pinned served config, Qwen2.5 Coder 1.5B base config comparison, Qwen2.5 lineage comparison, and a direct safetensors header range read. Bound weights use the exact 3.087428608 GB BF16 resident package; tie_word_embeddings is true and no lm_head.weight is stored, so the tied embedding/output table stays in ordinary text-decode traffic." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B-Instruct/raw/2e1fd397ee46e1388853d2af2c993145b0f1098a/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Coder-1.5B-Instruct" }, { "label": "Pinned base config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B/raw/df3ce67c0e24480f20468b6ef2894622d69eb73b/config.json" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B-Instruct/resolve/2e1fd397ee46e1388853d2af2c993145b0f1098a/model.safetensors" } ] }, { "id": "lmstudio-community/gemma-4-E4B-it-GGUF", "name": "gemma-4-E4B-it-GGUF GGUF", "short_name": "gemma 4 E4B it", "license": "apache-2.0", "hf_pipeline_tag": "unknown", "hf_downloads": 770548, "tags": [ "gguf", "base_model:google/gemma-4-E4B-it", "base_model:quantized:google/gemma-4-E4B-it", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 7.9962, "active_params_b": 7.9962, "layers": 42, "attention_heads": 8, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q4_K_M", "weight_bytes_per_param": 0.7096622095644453, "resident_weight_gb": 5.335289664, "active_weight_gb": 3.007356224, "kv_alloc_gb_per_1k_tokens": 0.086016, "kv_read_gb_per_1k_tokens": 0.086016, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, LM Studio model card, pinned Google Gemma 4 E4B IT config, Transformers Gemma 4 PLE documentation, GGUF linked-object HEAD checks, and a direct selected-file GGUF header range read. Bound weights use the selected Q4_K_M file size 5.335289664 GB for residency and 3.007356224 GB of ordinary swept tensor traffic after excluding the resident per-layer token embedding lookup table." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/gemma-4-E4B-it-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/google/gemma-4-E4B-it/raw/fee6332c1abaafb77f6f9624236c63aa2f1d0187/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-E4B-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-E4B-it-GGUF" }, { "label": "Selected GGUF header audit", "url": "https://huggingface.co/lmstudio-community/gemma-4-E4B-it-GGUF/resolve/53a691ddc52708042c56f80cdaf47f8a1daf051e/gemma-4-E4B-it-Q4_K_M.gguf" } ] }, { "id": "google/gemma-3-270m-it", "name": "gemma-3-270m-it BF16", "short_name": "gemma 3 270m it", "license": "gemma", "hf_pipeline_tag": "text-generation", "hf_downloads": 1035190, "tags": [ "transformers", "safetensors", "gemma3_text", "text-generation", "gemma3", "gemma", "google", "conversational", "arxiv:2503.19786", "arxiv:1905.07830", "arxiv:1905.10044", "arxiv:1911.11641", "arxiv:1705.03551", "arxiv:1911.01547", "arxiv:1907.10641", "arxiv:2311.07911", "arxiv:2311.12022", "arxiv:2411.04368", "arxiv:1904.09728", "arxiv:1903.00161", "arxiv:2009.03300", "arxiv:2304.06364", "arxiv:2103.03874", "arxiv:2110.14168", "arxiv:2108.07732", "arxiv:2107.03374", "arxiv:2403.07974", "arxiv:2305.03111", "arxiv:2405.04520", "arxiv:2210.03057", "arxiv:2106.03193", "arxiv:1910.11856", "arxiv:2502.12404", "arxiv:2502.21228", "arxiv:2404.16816", "arxiv:2104.12756", "arxiv:2311.16502", "arxiv:2203.10244", "arxiv:2404.12390", "arxiv:1810.12440", "arxiv:1908.02660", "arxiv:2310.02255", "arxiv:2312.11805", "base_model:google/gemma-3-270m", "base_model:finetune:google/gemma-3-270m", "license:gemma", "eval-results", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma3_text (gated config inaccessible)", "total_params_b": 0.268098176, "active_params_b": 0.268098176, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 0.536196352, "active_weight_gb": 0.536196352, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manually reviewed on 2026-07-05 from current Hugging Face API metadata and direct gated access checks. The repo is manually gated; unauthenticated raw config access returned 401 and hf download with the configured osolmaz CLI token returned access denied. Production bounds intentionally fail closed because max context and KV geometry cannot be verified without guessing." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-3-270m-it" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-3-270m" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-3-270m-it" }, { "label": "Pinned gated config access check", "url": "https://huggingface.co/google/gemma-3-270m-it/raw/ac82b4e820549b854eebf28ce6dedaf9fdfa17b3/config.json" } ] }, { "id": "Qwen/Qwen3.5-122B-A10B", "name": "Qwen3.5-122B-A10B BF16/F32", "short_name": "Qwen3.5 122B A10B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 784041, "tags": [ "transformers", "safetensors", "qwen3_5_moe", "image-text-to-text", "conversational", "license:apache-2.0", "eval-results", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_5_moe_text", "total_params_b": 125.086497008, "active_params_b": 9.008456448, "layers": 48, "attention_heads": 32, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "BF16/F32", "weight_bytes_per_param": 2, "resident_weight_gb": 250.17300784, "active_weight_gb": 18.016912896, "kv_alloc_gb_per_1k_tokens": 0.024576, "kv_read_gb_per_1k_tokens": 0.024576, "kv_alloc_fixed_gb": 0.154533888, "kv_read_fixed_gb": 0.154533888, "notes": "Manually audited on 2026-07-05 from HF API commit dc4d348443bc740c68e2d77492492c11606384d5, config.json, FP8 derivative config comparison, and range-read safetensors headers. The authoritative bounds profile uses exact stored bytes: 250.173007840 GB resident, 10.769155584 GB fixed text traffic, 0.905969664 GB per routed expert, 8 routed experts plus 1 shared expert, 12 full-attention layers, and a fixed DeltaNet recurrent-state charge of 0.154533888 GB per session/token." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B/raw/dc4d348443bc740c68e2d77492492c11606384d5/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-122B-A10B" }, { "label": "Safetensors index", "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B/resolve/dc4d348443bc740c68e2d77492492c11606384d5/model.safetensors.index.json" }, { "label": "FP8 derivative config comparison", "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B-FP8/raw/a099dee70ccfcd8d5dda56aaa0b60cb8ecadabc9/config.json" } ] }, { "id": "unsloth/Qwen3.6-35B-A3B-MTP-GGUF", "name": "Qwen3.6-35B-A3B-MTP-GGUF GGUF", "short_name": "Qwen3.6 35B A3B MTP", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 725918, "tags": [ "transformers", "gguf", "unsloth", "qwen", "qwen3_5_moe", "image-text-to-text", "base_model:Qwen/Qwen3.6-35B-A3B", "base_model:quantized:Qwen/Qwen3.6-35B-A3B", "license:apache-2.0", "endpoints_compatible", "region:us", "imatrix", "conversational" ], "architecture": { "type": "MoE transformer", "detail": "qwen35moe / qwen3_5_moe_text", "total_params_b": 35.505251456, "active_params_b": 3, "layers": 40, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "GGUF UD-Q4_K_XL", "weight_bytes_per_param": 0.6436699381307431, "resident_weight_gb": 22.853663008, "active_weight_gb": 2.7514496, "kv_alloc_gb_per_1k_tokens": 0.02048, "kv_read_gb_per_1k_tokens": 0.02048, "kv_alloc_fixed_gb": 0.06488064, "kv_read_fixed_gb": 0.06488064, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, model card MTP command, pinned Qwen base config, selected linked-object HEAD checks, and direct GGUF header range read. The selected MTP example artifact is Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf at 22.853663008 GB resident; ordinary single-session swept traffic is 2.751449600 GB before KV/state reads. Bounds Engine v1 treats the MTP draft block as resident-only for ordinary non-speculative text decode." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/Qwen3.6-35B-A3B-MTP-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/995ad96eacd98c81ed38be0c5b274b04031597b0/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3.6-35B-A3B-MTP-GGUF" }, { "label": "Selected GGUF header audit", "url": "https://huggingface.co/unsloth/Qwen3.6-35B-A3B-MTP-GGUF/resolve/5bc3e238d916f48a861bac2f8a1990a0e9b7e98d/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf" } ] }, { "id": "AxionML/Qwen3.5-9B-NVFP4", "name": "Qwen3.5-9B-NVFP4 NVFP4", "short_name": "Qwen3.5 9B NVFP4", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 748434, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "AxionML", "ModelOpt", "Qwen3.5", "quantized", "NVFP4", "nvfp4", "sglang", "conversational", "base_model:Qwen/Qwen3.5-9B", "base_model:quantized:Qwen/Qwen3.5-9B", "license:apache-2.0", "endpoints_compatible", "8-bit", "modelopt", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 9.653104368, "active_params_b": 7.936684544, "layers": 32, "attention_heads": 16, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "ModelOpt NVFP4", "weight_bytes_per_param": 0.9697254084428232, "resident_weight_gb": 9.360860576, "active_weight_gb": 5.928020928, "kv_alloc_gb_per_1k_tokens": 0.032768, "kv_read_gb_per_1k_tokens": 0.032768, "kv_alloc_fixed_gb": 0.051904512, "kv_read_fixed_gb": 0.051904512, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, pinned ModelOpt NVFP4 config and sidecar, Qwen3.5 9B base config comparison, direct safetensors header range read, and the Transformers qwen3_5 implementation. Bound weights separate the exact 9.360860576 GB resident package from 5.928020928 GB ordinary text-decode swept traffic; the audited profile supersedes the generated ideal 4-bit estimate and adds the fixed DeltaNet runtime state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/AxionML/Qwen3.5-9B-NVFP4" }, { "label": "Model config", "url": "https://huggingface.co/AxionML/Qwen3.5-9B-NVFP4/raw/97aef92393f126bf649f310cd40861be8dad3279/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-9B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/AxionML/Qwen3.5-9B-NVFP4" }, { "label": "Pinned base config", "url": "https://huggingface.co/Qwen/Qwen3.5-9B/raw/c202236235762e1c871ad0ccb60c8ee5ba337b9a/config.json" }, { "label": "ModelOpt quantization sidecar", "url": "https://huggingface.co/AxionML/Qwen3.5-9B-NVFP4/raw/97aef92393f126bf649f310cd40861be8dad3279/hf_quant_config.json" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/AxionML/Qwen3.5-9B-NVFP4/resolve/97aef92393f126bf649f310cd40861be8dad3279/model.safetensors" } ] }, { "id": "lmg-anon/vntl-llama3-8b-v2-gguf", "name": "VNTL Llama 3 8B V2 GGUF Q5_K_M", "short_name": "VNTL Llama 3 8B V2", "license": "llama3", "hf_pipeline_tag": "translation", "hf_downloads": 750593, "tags": [ "gguf", "translation", "ja", "en", "dataset:lmg-anon/VNTL-v5-1k", "base_model:rinna/llama-3-youko-8b", "base_model:quantized:rinna/llama-3-youko-8b", "license:llama3", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "llama (LlamaForCausalLM, Q5_K_M GGUF)", "total_params_b": 8.030261248, "active_params_b": 8.030261248, "layers": 32, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 8192, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q5_K_M", "weight_bytes_per_param": 0.7139233495582533, "resident_weight_gb": 5.732991008, "active_weight_gb": 5.363982336, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "notes": "Manually audited on 2026-07-05 from HF API commit ab7c8285d386fab7cd89e951f62e129085df6372, model card, rinna base-model API/config metadata, linked GGUF object checks, and a GGUF tensor-index range read. The authoritative bounds profile uses the selected vntl-llama3-8b-v2-hf-q5_k_m.gguf artifact: 5.732991008 GB resident, 5.363982336 GB swept text traffic, 0.369008672 GB resident-only token embedding plus GGUF metadata/header/file overhead, 8192-token context, and full-context FP16 K/V cache. The repo's Q8_0 artifact is not included in this row." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmg-anon/vntl-llama3-8b-v2-gguf" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmg-anon/vntl-llama3-8b-v2-gguf" }, { "label": "Base model config", "url": "https://huggingface.co/rinna/llama-3-youko-8b/raw/6af890bc6294b8f311e25a1462fcd38554d95b8e/config.json" }, { "label": "Base model API metadata", "url": "https://huggingface.co/api/models/rinna/llama-3-youko-8b" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/lmg-anon/vntl-llama3-8b-v2-gguf/resolve/ab7c8285d386fab7cd89e951f62e129085df6372/vntl-llama3-8b-v2-hf-q5_k_m.gguf" } ] }, { "id": "Qwen/Qwen2.5-32B-Instruct-AWQ", "name": "Qwen2.5-32B-Instruct-AWQ 4-bit", "short_name": "Qwen2.5 32B Instruct", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 787926, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "chat", "conversational", "en", "arxiv:2309.00071", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-32B-Instruct", "base_model:quantized:Qwen/Qwen2.5-32B-Instruct", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "4-bit", "awq", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM, AWQ GEMM)", "total_params_b": 32.763876352, "active_params_b": 32.763876352, "layers": 64, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "AWQ 4-bit GEMM", "weight_bytes_per_param": 0.5900474986395108, "resident_weight_gb": 19.328804864, "active_weight_gb": 17.771669504, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Manually audited on 2026-07-05 from current HF API metadata, model card, served AWQ config, current base-model API/config comparison, safetensors index, linked-object HEAD checks, and direct safetensors shard header range reads. Bound weights separate the exact 19.328804864 GB resident AWQ package from the 17.771669504 GB ordinary text-decode swept traffic; the audited profile supersedes the generated ideal 4-bit estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct-AWQ" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct-AWQ/raw/5c7cb76a268fc6cfbb9c4777eb24ba6e27f9ee6c/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-32B-Instruct-AWQ" }, { "label": "Base model API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-32B-Instruct" }, { "label": "Safetensors index", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct-AWQ/raw/5c7cb76a268fc6cfbb9c4777eb24ba6e27f9ee6c/model.safetensors.index.json" } ] }, { "id": "meta-llama/Llama-3.1-70B-Instruct", "name": "Llama-3.1-70B-Instruct BF16", "short_name": "Llama 3.1 70B", "license": "llama3.1", "hf_pipeline_tag": "text-generation", "hf_downloads": 1121582, "tags": [ "transformers", "safetensors", "llama", "text-generation", "facebook", "meta", "pytorch", "llama-3", "conversational", "en", "de", "fr", "it", "pt", "hi", "es", "th", "arxiv:2204.05149", "base_model:meta-llama/Llama-3.1-70B", "base_model:finetune:meta-llama/Llama-3.1-70B", "license:llama3.1", "eval-results", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gated config unavailable", "total_params_b": 70.553706496, "active_params_b": 70.553706496, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 141.107412992, "active_weight_gb": 141.107412992, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Fail-closed on 2026-07-05 from current Hugging Face API metadata and access checks. The API records 70.553706496B BF16 safetensors parameters at commit 1605565b47bb9346c5515c34102e054115b4f98b, but raw config, README, model.safetensors.index.json, and hf download config.json all require manual approval in this environment. KV geometry, max context, tied embeddings, and swept decode traffic are intentionally unknown until direct gated evidence is available." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/meta-llama/Llama-3.1-70B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.1-70B-Instruct" }, { "label": "Gated config access check", "url": "https://huggingface.co/meta-llama/Llama-3.1-70B-Instruct/raw/1605565b47bb9346c5515c34102e054115b4f98b/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/meta-llama/Llama-3.1-70B" } ] }, { "id": "meta-llama/Llama-4-Scout-17B-16E-Instruct", "name": "Llama-4-Scout-17B-16E-Instruct BF16", "short_name": "Llama 4 Scout 17B 16E", "license": "other", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 725196, "tags": [ "transformers", "safetensors", "llama4", "image-text-to-text", "facebook", "meta", "pytorch", "llama", "conversational", "ar", "de", "en", "es", "fr", "hi", "id", "it", "pt", "th", "tl", "vi", "arxiv:2204.05149", "base_model:meta-llama/Llama-4-Scout-17B-16E", "base_model:finetune:meta-llama/Llama-4-Scout-17B-16E", "license:other", "eval-results", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Unknown gated multimodal MoE", "detail": "gated config unavailable", "total_params_b": 108.641793536, "active_params_b": 108.641793536, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 217.283587072, "active_weight_gb": 217.283587072, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Fail-closed on 2026-07-05 from current Hugging Face API metadata and access checks. The API records 108.641793536B BF16 safetensors parameters at commit 92f3b1597a195b523d8d9e5700e57e4fbb8f20d3, but raw config, README, model.safetensors.index.json, and hf download config.json all require manual approval in this environment. Multimodal architecture, MoE routing, KV geometry, max context, active/swept traffic, and resident component split are intentionally unknown until direct gated evidence is available." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-4-Scout-17B-16E-Instruct" }, { "label": "Gated config access check", "url": "https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct/raw/92f3b1597a195b523d8d9e5700e57e4fbb8f20d3/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E" } ] }, { "id": "meta-llama/Llama-2-7b-hf", "name": "Llama-2-7b-hf F32/F16", "short_name": "Llama 2 7b", "license": "llama2", "hf_pipeline_tag": "text-generation", "hf_downloads": 679555, "tags": [ "transformers", "pytorch", "safetensors", "llama", "text-generation", "facebook", "meta", "llama-2", "en", "arxiv:2307.09288", "license:llama2", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Unknown gated model", "detail": "gated config unavailable", "total_params_b": 6.738417664, "active_params_b": 6.738417664, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "F32/F16", "weight_bytes_per_param": 2.000000607857839, "resident_weight_gb": 13.476839424, "active_weight_gb": 13.476839424, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Fail-closed on 2026-07-05 from current Hugging Face API metadata and access checks. The API records 6.738415616B F16 and 2048 F32 safetensors parameters at commit 01c7f73d771dfac7d292323805ebc428287df4f9, but raw config, README, model.safetensors.index.json, and hf download config.json all require manual approval in this environment. KV geometry, max context, tied embeddings, and swept decode traffic are intentionally unknown until direct gated evidence is available." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/meta-llama/Llama-2-7b-hf" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-2-7b-hf" }, { "label": "Gated config access check", "url": "https://huggingface.co/meta-llama/Llama-2-7b-hf/raw/01c7f73d771dfac7d292323805ebc428287df4f9/config.json" } ] }, { "id": "Qwen/Qwen3-30B-A3B-Instruct-2507", "name": "Qwen3-30B-A3B-Instruct-2507 BF16", "short_name": "Qwen3 30B A3B Instruct 2507", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 1092820, "tags": [ "transformers", "safetensors", "qwen3_moe", "text-generation", "conversational", "arxiv:2402.17463", "arxiv:2407.02490", "arxiv:2501.15383", "arxiv:2404.06654", "arxiv:2505.09388", "license:apache-2.0", "eval-results", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_moe (Qwen3MoeForCausalLM)", "total_params_b": 30.532122624, "active_params_b": 3.3, "layers": 48, "attention_heads": 32, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 61.064245248, "active_weight_gb": 6.083735552, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, model card, pinned served config, safetensors index, and direct shard header range reads. Bound weights separate the exact 61.064245248 GB resident BF16 package from the 6.083735552 GB ordinary single-session MoE text-decode traffic; the audited profile supersedes the generated rounded active-weight estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507/raw/0d7cf23991f47feeb3a57ecb4c9cee8ea4a17bfe/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-30B-A3B-Instruct-2507" }, { "label": "Safetensors index", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507/raw/0d7cf23991f47feeb3a57ecb4c9cee8ea4a17bfe/model.safetensors.index.json" } ] }, { "id": "meta-llama/Llama-3.3-70B-Instruct", "name": "Llama-3.3-70B-Instruct BF16", "short_name": "Llama 3.3 70B", "license": "llama3.3", "hf_pipeline_tag": "text-generation", "hf_downloads": 740835, "tags": [ "transformers", "safetensors", "llama", "text-generation", "facebook", "meta", "pytorch", "llama-3", "conversational", "en", "fr", "it", "pt", "hi", "es", "th", "de", "arxiv:2204.05149", "base_model:meta-llama/Llama-3.1-70B", "base_model:finetune:meta-llama/Llama-3.1-70B", "license:llama3.3", "eval-results", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "unknown gated LlamaForCausalLM", "total_params_b": 70.553706496, "active_params_b": 70.553706496, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 141.107412992, "active_weight_gb": 141.107412992, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Fail-closed on 2026-07-05 from current Hugging Face API metadata and access checks. The API records 70.553706496B BF16 safetensors parameters at commit 6f6073b423013f6a7d4d9f39144961bfbfbc386b, but raw config, README, model.safetensors.index.json, and hf download config.json all require manual approval in this environment. KV geometry, max context, tied embeddings, and swept decode traffic are intentionally unknown until direct gated evidence is available." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.3-70B-Instruct" }, { "label": "Gated config access check", "url": "https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct/raw/6f6073b423013f6a7d4d9f39144961bfbfbc386b/config.json" } ] }, { "id": "lmstudio-community/Qwen3.5-9B-MLX-4bit", "name": "Qwen3.5-9B-MLX-4bit 4-bit", "short_name": "Qwen3.5 9B MLX 4bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 762722, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "mlx", "conversational", "base_model:Qwen/Qwen3.5-9B", "base_model:quantized:Qwen/Qwen3.5-9B", "license:apache-2.0", "endpoints_compatible", "4-bit", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 9.409813744, "active_params_b": 7.936684544, "layers": 32, "attention_heads": 16, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "MLX 4-bit affine", "weight_bytes_per_param": 0.6323252215054683, "resident_weight_gb": 5.95006256, "active_weight_gb": 4.46591232, "kv_alloc_gb_per_1k_tokens": 0.032768, "kv_read_gb_per_1k_tokens": 0.032768, "kv_alloc_fixed_gb": 0.051904512, "kv_read_fixed_gb": 0.051904512, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, pinned served MLX config, Qwen3.5 9B base config comparison, model card, safetensors index, direct shard header range reads, and the Transformers qwen3_5 implementation. Bound weights separate the exact 5.95006256 GB resident MLX package from the 4.46591232 GB ordinary text-decode swept traffic; the audited profile supersedes the generated ideal 4-bit estimate and adds the fixed DeltaNet runtime state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/Qwen3.5-9B-MLX-4bit" }, { "label": "Model config", "url": "https://huggingface.co/lmstudio-community/Qwen3.5-9B-MLX-4bit/raw/b455506b0f574c74616dbcd56879bde38fafcff3/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-9B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3.5-9B-MLX-4bit" }, { "label": "Pinned base config", "url": "https://huggingface.co/Qwen/Qwen3.5-9B/raw/c202236235762e1c871ad0ccb60c8ee5ba337b9a/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/lmstudio-community/Qwen3.5-9B-MLX-4bit/raw/b455506b0f574c74616dbcd56879bde38fafcff3/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen3-1.7B-Base", "name": "Qwen3-1.7B-Base BF16", "short_name": "Qwen3 1.7B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 670780, "tags": [ "transformers", "safetensors", "qwen3", "text-generation", "conversational", "arxiv:2505.09388", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 1.720574976, "active_params_b": 1.720574976, "layers": 28, "attention_heads": 16, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 3.441149952, "active_weight_gb": 3.441149952, "kv_alloc_gb_per_1k_tokens": 0.114688, "kv_read_gb_per_1k_tokens": 0.114688, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, pinned served config, and a direct model.safetensors header range read. The checkpoint stores 3.441149952 GB of BF16 tensors, has no separate lm_head.weight tensor, and keeps model.embed_tokens.weight swept as the tied output projection." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-1.7B-Base" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-1.7B-Base/raw/ea980cb0a6c2ae4b936e82123acc929f1cec04c1/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-1.7B-Base" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/Qwen/Qwen3-1.7B-Base/resolve/ea980cb0a6c2ae4b936e82123acc929f1cec04c1/model.safetensors" } ] }, { "id": "moonshotai/Kimi-K2.7-Code", "name": "Kimi-K2.7-Code 4-bit", "short_name": "Kimi K2.7 Code", "license": "other", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 850031, "tags": [ "transformers", "safetensors", "kimi_k25", "feature-extraction", "compressed-tensors", "image-text-to-text", "conversational", "custom_code", "license:other", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "kimi_k2 (DeepseekV3ForCausalLM)", "total_params_b": 1058.589420528, "active_params_b": 32, "layers": 61, "attention_heads": 64, "kv_heads": 64, "head_dim": 192, "max_context_tokens": 262144, "routed_experts": 384, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "INT4/BF16 compressed-tensors", "weight_bytes_per_param": 0.562208710190164, "resident_weight_gb": 595.148192736, "active_weight_gb": 32.986516736, "kv_alloc_gb_per_1k_tokens": 2.49856, "kv_read_gb_per_1k_tokens": 2.49856, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, pinned served config, model card, deployment guide, direct safetensors shard header range reads, and the shipped custom Kimi/DeepSeek runtime. Bound weights use exact stored INT4/BF16/F32 bytes, separate 3.29109808 GB resident-only multimodal/input-embedding bytes, and charge expanded BF16 K/V cache traffic for ordinary text decode." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/moonshotai/Kimi-K2.7-Code" }, { "label": "Model config", "url": "https://huggingface.co/moonshotai/Kimi-K2.7-Code/raw/74797c9c62378b951a1f6fcf5c4631024e9b8bef/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/moonshotai/Kimi-K2.7-Code" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/moonshotai/Kimi-K2.7-Code/resolve/74797c9c62378b951a1f6fcf5c4631024e9b8bef/model.safetensors.index.json" }, { "label": "Deployment guidance", "url": "https://huggingface.co/moonshotai/Kimi-K2.7-Code/raw/74797c9c62378b951a1f6fcf5c4631024e9b8bef/docs/deploy_guidance.md" } ] }, { "id": "meta-llama/Llama-3.2-3B", "name": "Llama-3.2-3B BF16", "short_name": "Llama 3.2 3B", "license": "llama3.2", "hf_pipeline_tag": "text-generation", "hf_downloads": 640543, "tags": [ "transformers", "safetensors", "llama", "text-generation", "facebook", "meta", "pytorch", "llama-3", "en", "de", "fr", "it", "pt", "hi", "es", "th", "arxiv:2204.05149", "arxiv:2405.16406", "license:llama3.2", "eval-results", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gated config unavailable", "total_params_b": 3.212749824, "active_params_b": 3.212749824, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 6.425499648, "active_weight_gb": 6.425499648, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manually checked on 2026-07-05 from current Hugging Face API metadata and direct raw-file/HF CLI access attempts. The repo is gated and the configured osolmaz token lacks approval, so config, tensor index, max context, KV geometry, tied embeddings, and swept decode traffic are intentionally left unverified. Bounds fail closed through the unsupported model profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/meta-llama/Llama-3.2-3B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.2-3B" }, { "label": "Gated config access check", "url": "https://huggingface.co/meta-llama/Llama-3.2-3B/raw/13afe5124825b4f3751f836b40dafda64c1ed062/config.json" } ] }, { "id": "nvidia/Kimi-K2.6-NVFP4", "name": "Kimi-K2.6-NVFP4 NVFP4", "short_name": "Kimi K2.6 NVFP4", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 768850, "tags": [ "Model Optimizer", "safetensors", "kimi_k25", "nvidia", "ModelOpt", "Kimi-K2.6", "quantized", "FP4", "fp4", "text-generation", "conversational", "custom_code", "base_model:moonshotai/Kimi-K2.6", "base_model:quantized:moonshotai/Kimi-K2.6", "license:other", "modelopt", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "kimi_k25 (KimiK25ForConditionalGeneration / DeepseekV3ForCausalLM)", "total_params_b": 1058.589420528, "active_params_b": 32, "layers": 61, "attention_heads": 64, "kv_heads": 64, "head_dim": 192, "max_context_tokens": 262144, "routed_experts": 384, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "NVFP4/BF16/FP8 ModelOpt", "weight_bytes_per_param": 0.5, "resident_weight_gb": 595.148146656, "active_weight_gb": 32.986470656, "kv_alloc_gb_per_1k_tokens": 1.24928, "kv_read_gb_per_1k_tokens": 1.24928, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, NVIDIA model card, served config, hf_quant_config, custom runtime review, safetensors index, and direct shard header range reads. Bound weights separate the exact 595.148146656 GB resident ModelOpt package from 32.986470656 GB single-session ordinary text-decode MoE traffic and charge expanded FP8 K/V cache streams for ordinary text decode." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/Kimi-K2.6-NVFP4" }, { "label": "Model config", "url": "https://huggingface.co/nvidia/Kimi-K2.6-NVFP4/raw/2fd3a800dedd098b8327eb49e93ebc75f85da19f/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/Kimi-K2.6-NVFP4" }, { "label": "ModelOpt quantization config", "url": "https://huggingface.co/nvidia/Kimi-K2.6-NVFP4/raw/2fd3a800dedd098b8327eb49e93ebc75f85da19f/hf_quant_config.json" }, { "label": "Safetensors index", "url": "https://huggingface.co/nvidia/Kimi-K2.6-NVFP4/resolve/2fd3a800dedd098b8327eb49e93ebc75f85da19f/model.safetensors.index.json" }, { "label": "Base model card", "url": "https://huggingface.co/moonshotai/Kimi-K2.6" } ] }, { "id": "Qwen/Qwen2-7B-Instruct", "name": "Qwen2-7B-Instruct BF16", "short_name": "Qwen2 7B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 692373, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "chat", "conversational", "en", "arxiv:2309.00071", "base_model:Qwen/Qwen2-7B", "base_model:finetune:Qwen/Qwen2-7B", "license:apache-2.0", "eval-results", "text-generation-inference", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 7.615616512, "active_params_b": 7.070619136, "layers": 28, "attention_heads": 28, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 15.231233024, "active_weight_gb": 14.141238272, "kv_alloc_gb_per_1k_tokens": 0.057344, "kv_read_gb_per_1k_tokens": 0.057344, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, pinned served config, base config comparison, and direct safetensors shard header range reads. Bound weights separate the exact 15.231233024 GB resident BF16 package from 14.141238272 GB ordinary text-decode swept traffic; the served Instruct config records 32768 max position embeddings while the base config records 131072." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2-7B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2-7B-Instruct/raw/f2826a00ceef68f0f2b946d945ecc0477ce4450c/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2-7B-Instruct" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Qwen/Qwen2-7B-Instruct/resolve/f2826a00ceef68f0f2b946d945ecc0477ce4450c/model.safetensors.index.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2-7B" } ] }, { "id": "unsloth/gemma-4-E2B-it-GGUF", "name": "gemma-4-E2B-it-GGUF GGUF", "short_name": "gemma 4 E2B it", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 608014, "tags": [ "gguf", "gemma4", "unsloth", "gemma", "google", "image-text-to-text", "base_model:google/gemma-4-E2B-it", "base_model:quantized:google/gemma-4-E2B-it", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 4.647450147, "active_params_b": 2.298639907, "layers": 35, "attention_heads": 8, "kv_heads": 1, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 9.311303552, "active_weight_gb": 4.597865568, "kv_alloc_gb_per_1k_tokens": 0.03584, "kv_read_gb_per_1k_tokens": 0.03584, "notes": "Manually audited on 2026-07-05 from current Hugging Face API GGUF metadata, selected-artifact HEAD checks, Google Gemma 4 E2B config, Transformers PLE documentation, and a direct GGUF header/tensor-index range read. Bound weights use the selected gemma-4-E2B-it-BF16.gguf linked size 9.311303552 GB and PLE-aware ordinary text-decode swept traffic 4.597865568 GB; mmproj and MTP sidecars are excluded unless explicitly loaded." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/gemma-4-E2B-it-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/gemma-4-E2B-it-GGUF" }, { "label": "Selected GGUF tensor index", "url": "https://huggingface.co/unsloth/gemma-4-E2B-it-GGUF/resolve/ecc8b33b2c50598815e4b0f7cea6088e3ae7adb8/gemma-4-E2B-it-BF16.gguf" }, { "label": "Base model config", "url": "https://huggingface.co/google/gemma-4-E2B-it/raw/70af34e20bd4b7a91f0de6b22675850c43922a03/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-E2B-it" } ] }, { "id": "unsloth/gemma-4-E4B-it-GGUF", "name": "gemma-4-E4B-it-GGUF GGUF", "short_name": "gemma 4 E4B it", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 677461, "tags": [ "gguf", "gemma4", "unsloth", "gemma", "google", "image-text-to-text", "base_model:google/gemma-4-E4B-it", "base_model:quantized:google/gemma-4-E4B-it", "license:apache-2.0", "endpoints_compatible", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 7.9962, "active_params_b": 7.9962, "layers": 42, "attention_heads": 8, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF", "weight_bytes_per_param": 2, "resident_weight_gb": 15.9924, "active_weight_gb": 15.9924, "kv_alloc_gb_per_1k_tokens": 0.086016, "kv_read_gb_per_1k_tokens": 0.086016, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=677461; data quality: config, params, base:google/gemma-4-E4B-it. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/gemma-4-E4B-it-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/unsloth/gemma-4-E4B-it-GGUF/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-E4B-it" } ] }, { "id": "Qwen/Qwen2-Audio-7B-Instruct", "name": "Qwen2-Audio-7B-Instruct BF16", "short_name": "Qwen2 Audio 7B", "license": "apache-2.0", "hf_pipeline_tag": "audio-text-to-text", "hf_downloads": 672631, "tags": [ "transformers", "safetensors", "qwen2_audio", "text2text-generation", "chat", "audio", "audio-text-to-text", "en", "arxiv:2407.10759", "arxiv:2311.07919", "license:apache-2.0", "endpoints_compatible" ], "architecture": { "type": "Dense transformer", "detail": "qwen2", "total_params_b": 8.3971, "active_params_b": 8.3971, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": 8192, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 16.7942, "active_weight_gb": 16.7942, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=672631; data quality: config, params. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2-Audio-7B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2-Audio-7B-Instruct/raw/main/config.json" } ] }, { "id": "Qwen/Qwen3-VL-8B-Instruct-FP8", "name": "Qwen3-VL-8B-Instruct-FP8 FP8", "short_name": "Qwen3 VL 8B Instruct FP8", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 664215, "tags": [ "transformers", "safetensors", "qwen3_vl", "image-text-to-text", "conversational", "arxiv:2505.09388", "arxiv:2502.13923", "arxiv:2409.12191", "arxiv:2308.12966", "base_model:Qwen/Qwen3-VL-8B-Instruct", "base_model:quantized:Qwen/Qwen3-VL-8B-Instruct", "license:apache-2.0", "endpoints_compatible", "fp8", "deploy:azure" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_vl_text", "total_params_b": 8.7675, "active_params_b": 8.7675, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "FP8", "weight_bytes_per_param": 1, "resident_weight_gb": 8.7675, "active_weight_gb": 8.7675, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=664215; data quality: config, params, base:Qwen/Qwen3-VL-8B-Instruct. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct-FP8" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct-FP8/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct" } ] }, { "id": "nvidia/Cosmos-Reason2-2B", "name": "Cosmos-Reason2-2B BF16", "short_name": "Cosmos Reason2 2B", "license": "other", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 663643, "tags": [ "cosmos", "safetensors", "qwen3_vl", "nvidia", "conversational", "image-text-to-text", "arxiv:2503.06800", "arxiv:2406.10721", "arxiv:2603.18178", "arxiv:2312.14115", "base_model:Qwen/Qwen3-VL-2B-Instruct", "base_model:finetune:Qwen/Qwen3-VL-2B-Instruct", "license:other" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_vl_text", "total_params_b": 2.4387, "active_params_b": 2.4387, "layers": 28, "attention_heads": 16, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 4.8774, "active_weight_gb": 4.8774, "kv_alloc_gb_per_1k_tokens": 0.114688, "kv_read_gb_per_1k_tokens": 0.114688, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=663643; data quality: config, params, base:Qwen/Qwen3-VL-2B-Instruct. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/Cosmos-Reason2-2B" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3-VL-2B-Instruct/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-2B-Instruct" } ] }, { "id": "microsoft/Florence-2-large", "name": "Florence-2-large F16", "short_name": "Florence 2 large", "license": "mit", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 660679, "tags": [ "transformers", "pytorch", "safetensors", "florence2", "image-text-to-text", "vision", "custom_code", "arxiv:2311.06242", "license:mit", "endpoints_compatible" ], "architecture": { "type": "Dense transformer", "detail": "unknown", "total_params_b": 0.7767, "active_params_b": 0.7767, "layers": 12, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": 4096, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "F16", "weight_bytes_per_param": 2, "resident_weight_gb": 1.5534, "active_weight_gb": 1.5534, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=660679; data quality: config, params. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/microsoft/Florence-2-large" }, { "label": "Model config", "url": "https://huggingface.co/microsoft/Florence-2-large/raw/main/config.json" } ] }, { "id": "Qwen/Qwen3-4B-Base", "name": "Qwen3-4B-Base BF16", "short_name": "Qwen3 4B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 648944, "tags": [ "transformers", "safetensors", "qwen3", "text-generation", "conversational", "arxiv:2505.09388", "license:apache-2.0", "text-generation-inference", "endpoints_compatible" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 4.0225, "active_params_b": 4.0225, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 8.045, "active_weight_gb": 8.045, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=648944; data quality: config, params. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-4B-Base" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-4B-Base/raw/main/config.json" } ] }, { "id": "cyankiwi/Hermes-4-14B-AWQ-4bit", "name": "Hermes-4-14B-AWQ-4bit 4-bit", "short_name": "Hermes 4 14B AWQ 4bit", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 647857, "tags": [ "transformers", "safetensors", "qwen3", "text-generation", "Qwen-3-14B", "instruct", "finetune", "reasoning", "hybrid-mode", "chatml", "function calling", "tool use", "json mode", "structured outputs", "atropos", "dataforge", "long context", "roleplaying", "chat", "conversational", "en", "arxiv:2508.18255", "base_model:NousResearch/Hermes-4-14B", "base_model:quantized:NousResearch/Hermes-4-14B", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "compressed-tensors" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 3.6206, "active_params_b": 3.6206, "layers": 40, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "4-bit", "weight_bytes_per_param": 0.5, "resident_weight_gb": 1.8103, "active_weight_gb": 1.8103, "kv_alloc_gb_per_1k_tokens": 0.16384, "kv_read_gb_per_1k_tokens": 0.16384, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=647857; data quality: config, params, base:NousResearch/Hermes-4-14B. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/cyankiwi/Hermes-4-14B-AWQ-4bit" }, { "label": "Model config", "url": "https://huggingface.co/cyankiwi/Hermes-4-14B-AWQ-4bit/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/NousResearch/Hermes-4-14B" } ] }, { "id": "Qwen/Qwen2.5-Omni-7B", "name": "Qwen2.5-Omni-7B F32/BF16", "short_name": "Qwen2.5 Omni 7B", "license": "other", "hf_pipeline_tag": "any-to-any", "hf_downloads": 647792, "tags": [ "transformers", "safetensors", "qwen2_5_omni", "multimodal", "any-to-any", "en", "arxiv:2503.20215", "license:other", "endpoints_compatible" ], "architecture": { "type": "Dense transformer", "detail": "qwen2_5_omni (Qwen2_5OmniModel)", "total_params_b": 10.7322, "active_params_b": 10.7322, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": 2048, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "F32/BF16", "weight_bytes_per_param": 2.0837, "resident_weight_gb": 22.3627, "active_weight_gb": 22.3627, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Generated from Hugging Face metadata on 2026-07-01; downloads=647792; data quality: config, params. Bound weights use recorded repo precision when detectable, otherwise metadata-estimated precision." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Omni-7B" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-Omni-7B/raw/main/config.json" } ] }, { "id": "allenai/Molmo2-8B", "name": "Molmo2-8B F32", "short_name": "Molmo2 8B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 603471, "tags": [ "transformers", "safetensors", "molmo2", "image-text-to-text", "multimodal", "olmo", "molmo", "conversational", "custom_code", "en", "dataset:allenai/Molmo2-Cap", "dataset:allenai/Molmo2-VideoCapQA", "dataset:allenai/Molmo2-VideoSubtitleQA", "dataset:allenai/Molmo2-AskModelAnything", "dataset:allenai/Molmo2-VideoPoint", "dataset:allenai/Molmo2-VideoTrack", "dataset:allenai/Molmo2-MultiImageQA", "dataset:allenai/Molmo2-SynMultiImageQA", "dataset:allenai/Molmo2-MultiImagePoint", "base_model:Qwen/Qwen3-8B", "base_model:finetune:Qwen/Qwen3-8B", "license:apache-2.0", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "molmo2_text", "total_params_b": 8.66170312, "active_params_b": 7.568405504, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 36864, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "F32", "weight_bytes_per_param": 4, "resident_weight_gb": 34.64681248, "active_weight_gb": 30.273622016, "kv_alloc_gb_per_1k_tokens": 0.294912, "kv_read_gb_per_1k_tokens": 0.294912, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, pinned served Molmo2 config, pinned custom modeling code, safetensors index, and direct range-read shard headers. Bound weights separate the 34.64681248 GB resident F32 multimodal package from the 30.273622016 GB ordinary text-decode swept subset; model.transformer.wte and model.vision_backbone are resident-only. The audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/allenai/Molmo2-8B" }, { "label": "Model config", "url": "https://huggingface.co/allenai/Molmo2-8B/raw/e28fa28597e5ec5e0cca2201dd8ab33d48bc4a1b/config.json" }, { "label": "Safetensors index", "url": "https://huggingface.co/allenai/Molmo2-8B/raw/e28fa28597e5ec5e0cca2201dd8ab33d48bc4a1b/model.safetensors.index.json" }, { "label": "Custom modeling code", "url": "https://huggingface.co/allenai/Molmo2-8B/raw/e28fa28597e5ec5e0cca2201dd8ab33d48bc4a1b/modeling_molmo2.py" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-8B" } ] }, { "id": "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B", "name": "DeepSeek-R1-Distill-Qwen-1.5B BF16", "short_name": "DeepSeek R1 Distill Qwen 1.5B", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 630704, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "conversational", "arxiv:2501.12948", "license:mit", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 1.777088, "active_params_b": 1.543714304, "layers": 28, "attention_heads": 12, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 3.554176, "active_weight_gb": 3.087428608, "kv_alloc_gb_per_1k_tokens": 0.028672, "kv_read_gb_per_1k_tokens": 0.028672, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, model card base-model statement, pinned served config, Qwen2.5-Math-1.5B base config comparison, and a direct safetensors header range read. Bound weights separate the 3.554176 GB resident BF16 package from the 3.087428608 GB ordinary text-decode swept subset; model.embed_tokens.weight is resident-only. The audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B" }, { "label": "Model config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B/raw/ad9f0ae0864d7fbcd1cd905e3c6c5b069cc8b562/config.json" }, { "label": "Safetensors file", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B/resolve/ad9f0ae0864d7fbcd1cd905e3c6c5b069cc8b562/model.safetensors" }, { "label": "Base config", "url": "https://huggingface.co/Qwen/Qwen2.5-Math-1.5B/raw/main/config.json" } ] }, { "id": "cyankiwi/GLM-4.5-Air-AWQ-4bit", "name": "GLM-4.5-Air-AWQ-4bit 4-bit", "short_name": "GLM 4.5 Air AWQ 4bit", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 511138, "tags": [ "transformers", "safetensors", "glm4_moe", "text-generation", "conversational", "en", "zh", "arxiv:2508.06471", "base_model:zai-org/GLM-4.5-Air", "base_model:quantized:zai-org/GLM-4.5-Air", "license:mit", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "glm4_moe (Glm4MoeForCausalLM)", "total_params_b": 110.4688, "active_params_b": 6.9043, "layers": 46, "attention_heads": 96, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "4-bit", "weight_bytes_per_param": 0.5, "resident_weight_gb": 63.394045248, "active_weight_gb": 9.598953664, "kv_alloc_gb_per_1k_tokens": 0.188416, "kv_read_gb_per_1k_tokens": 0.188416, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, model card, served compressed-tensors AWQ config, GLM-4.5-Air base config comparison, direct safetensors header range reads, and the Transformers glm4_moe implementation. Bound weights separate the 63.394045248 GB resident package from 6.095381824 GB fixed ordinary text traffic plus 0.43794648 GB per routed expert index; model.embed_tokens.weight is resident-only. The audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/cyankiwi/GLM-4.5-Air-AWQ-4bit" }, { "label": "Model config", "url": "https://huggingface.co/cyankiwi/GLM-4.5-Air-AWQ-4bit/raw/982ca6089b239302649a166da300ba83017391ff/config.json" }, { "label": "Safetensors index", "url": "https://huggingface.co/cyankiwi/GLM-4.5-Air-AWQ-4bit/raw/982ca6089b239302649a166da300ba83017391ff/model.safetensors.index.json" }, { "label": "Base model card", "url": "https://huggingface.co/zai-org/GLM-4.5-Air" } ] }, { "id": "MiniMaxAI/MiniMax-M2.5", "name": "MiniMax-M2.5 FP8", "short_name": "MiniMax M2.5", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 698560, "tags": [ "transformers", "safetensors", "minimax_m2", "text-generation", "conversational", "custom_code", "license:other", "eval-results", "endpoints_compatible", "fp8", "deploy:azure", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "minimax_m2 (MiniMaxM2ForCausalLM)", "total_params_b": 228.703644928, "active_params_b": 10.41655168, "layers": 62, "attention_heads": 48, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 196608, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "FP8", "weight_bytes_per_param": 1, "resident_weight_gb": 230.12163072, "active_weight_gb": 11.180086272, "kv_alloc_gb_per_1k_tokens": 0.253952, "kv_read_gb_per_1k_tokens": 0.253952, "notes": "Manually corrected on 2026-07-05 from current Hugging Face API metadata, served config, custom configuration/model code, safetensors index, and direct safetensors header range reads. Active traffic is fixed decode traffic plus 8 routed expert groups; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M2.5" }, { "label": "Model config", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M2.5/raw/f710177d938eff80b684d42c5aa84b382612f21f/config.json" }, { "label": "Safetensors index", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M2.5/raw/f710177d938eff80b684d42c5aa84b382612f21f/model.safetensors.index.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MiniMaxAI/MiniMax-M2.5" } ] }, { "id": "Qwen/Qwen2.5-72B-Instruct", "name": "Qwen2.5-72B-Instruct BF16", "short_name": "Qwen2.5 72B", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 627874, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "chat", "conversational", "en", "arxiv:2309.00071", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-72B", "base_model:finetune:Qwen/Qwen2.5-72B", "license:other", "text-generation-inference", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 72.706203648, "active_params_b": 71.46049536, "layers": 80, "attention_heads": 64, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 145.412407296, "active_weight_gb": 142.92099072, "kv_alloc_gb_per_1k_tokens": 0.32768, "kv_read_gb_per_1k_tokens": 0.32768, "notes": "Manually corrected on 2026-07-05 from current Hugging Face API metadata, served config, model card, base config comparison, safetensors index, and direct safetensors header range reads. Active traffic excludes only the resident input embedding lookup and includes the separate lm_head output projection; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-72B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-72B-Instruct/raw/495f39366efef23836d0cfae4fbe635880d2be31/config.json" }, { "label": "Safetensors index", "url": "https://huggingface.co/Qwen/Qwen2.5-72B-Instruct/raw/495f39366efef23836d0cfae4fbe635880d2be31/model.safetensors.index.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-72B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-72B-Instruct" } ] }, { "id": "lmstudio-community/Qwen3.5-9B-MLX-8bit", "name": "Qwen3.5-9B-MLX-8bit 8-bit", "short_name": "Qwen3.5 9B MLX 8bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 668158, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "mlx", "conversational", "base_model:Qwen/Qwen3.5-9B", "base_model:quantized:Qwen/Qwen3.5-9B", "license:apache-2.0", "endpoints_compatible", "8-bit", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 9.409813744, "active_params_b": 7.936684544, "layers": 32, "attention_heads": 16, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "MLX 8-bit affine", "weight_bytes_per_param": 1.1080382447153356, "resident_weight_gb": 10.426433504, "active_weight_gb": 8.433723904, "kv_alloc_gb_per_1k_tokens": 0.032768, "kv_read_gb_per_1k_tokens": 0.032768, "kv_alloc_fixed_gb": 0.051904512, "kv_read_fixed_gb": 0.051904512, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, pinned served MLX config, Qwen3.5 9B base config comparison, model card, safetensors index, direct shard header range reads, and the Transformers qwen3_5 implementation. Bound weights separate the exact 10.426433504 GB resident MLX package from the 8.433723904 GB ordinary text-decode swept traffic; the audited profile supersedes the generated ideal 8-bit estimate and adds the fixed DeltaNet runtime state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/Qwen3.5-9B-MLX-8bit" }, { "label": "Model config", "url": "https://huggingface.co/lmstudio-community/Qwen3.5-9B-MLX-8bit/raw/950e0459b483aec4324437bc5d45cf446b2d67bd/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-9B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3.5-9B-MLX-8bit" }, { "label": "Pinned base config", "url": "https://huggingface.co/Qwen/Qwen3.5-9B/raw/c202236235762e1c871ad0ccb60c8ee5ba337b9a/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/lmstudio-community/Qwen3.5-9B-MLX-8bit/raw/950e0459b483aec4324437bc5d45cf446b2d67bd/model.safetensors.index.json" } ] }, { "id": "lmstudio-community/gemma-4-12B-it-QAT-GGUF", "name": "Gemma 4 12B IT QAT GGUF Q4_0", "short_name": "Gemma 4 12B QAT GGUF", "license": "apache-2.0", "hf_pipeline_tag": "unknown", "hf_downloads": 711351, "tags": [ "gguf", "base_model:google/gemma-4-12B-it-qat-q4_0-unquantized", "base_model:quantized:google/gemma-4-12B-it-qat-q4_0-unquantized", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_unified_text", "total_params_b": 11.907350576, "active_params_b": 11.907350576, "layers": 48, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q4_0 QAT", "weight_bytes_per_param": 0.5858464076853765, "resident_weight_gb": 6.97587856, "active_weight_gb": 6.960055808, "kv_alloc_gb_per_1k_tokens": 0.335872, "kv_read_gb_per_1k_tokens": 0.335872, "notes": "Manually audited on 2026-07-05 from HF API commit 291406f49e16eff811c85ad8884d375f34138663, Google QAT unquantized base config, model card, linked GGUF file sizes, and a GGUF tensor-index range read. The authoritative bounds profile uses the selected gemma-4-12B-it-QAT-Q4_0.gguf artifact: 6.97587856 GB resident, 6.960055808 GB swept text traffic, 0.015822752 GB resident-only GGUF metadata/header bytes, and layered Gemma 4 hybrid KV. The separate mmproj GGUF file is not included unless explicitly loaded. Legacy per-1k KV fields show the short-context unsaturated coefficient; long-context calculation must use the profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/gemma-4-12B-it-QAT-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/google/gemma-4-12B-it-qat-q4_0-unquantized/raw/58540658b6c08edab2ddc1fbde7f28cc9987ced3/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-12B-it-qat-q4_0-unquantized" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-12B-it-QAT-GGUF" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/lmstudio-community/gemma-4-12B-it-QAT-GGUF/resolve/291406f49e16eff811c85ad8884d375f34138663/gemma-4-12B-it-QAT-Q4_0.gguf" } ] }, { "id": "lmstudio-community/Qwen3.6-35B-A3B-MLX-4bit", "name": "Qwen3.6-35B-A3B-MLX-4bit 4-bit", "short_name": "Qwen3.6 35B A3B MLX 4bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 648427, "tags": [ "transformers", "safetensors", "qwen3_5_moe", "image-text-to-text", "mlx", "conversational", "base_model:Qwen/Qwen3.6-35B-A3B", "base_model:quantized:Qwen/Qwen3.6-35B-A3B", "license:apache-2.0", "endpoints_compatible", "4-bit", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_5_moe_text", "total_params_b": 35.9518, "active_params_b": 3, "layers": 40, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "MLX 4-bit affine", "weight_bytes_per_param": 0.5674797108334115, "resident_weight_gb": 20.401929952, "active_weight_gb": 1.669560576, "kv_alloc_gb_per_1k_tokens": 0.02048, "kv_read_gb_per_1k_tokens": 0.02048, "kv_alloc_fixed_gb": 0.06488064, "kv_read_fixed_gb": 0.06488064, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, pinned served MLX config, Qwen3.6 35B A3B base config comparison, model card, safetensors index, direct shard header range reads, and the Transformers qwen3_5 implementation. Bound weights separate the exact 20.401929952 GB resident MLX package from the 1.669560576 GB single-session active ordinary text-decode traffic; the audited profile supersedes the generated ideal 4-bit estimate and adds the fixed DeltaNet runtime state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-35B-A3B-MLX-4bit" }, { "label": "Model config", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-35B-A3B-MLX-4bit/raw/0c4a20a6437ae5985ddc9eb1a3f122ee6c151c3b/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3.6-35B-A3B-MLX-4bit" }, { "label": "Pinned base config", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/995ad96eacd98c81ed38be0c5b274b04031597b0/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-35B-A3B-MLX-4bit/raw/0c4a20a6437ae5985ddc9eb1a3f122ee6c151c3b/model.safetensors.index.json" } ] }, { "id": "google/gemma-4-31B-it-assistant", "name": "Gemma 4 31B IT Assistant Drafter BF16", "short_name": "Gemma 4 31B IT Assistant", "license": "apache-2.0", "hf_pipeline_tag": "any-to-any", "hf_downloads": 806036, "tags": [ "transformers", "safetensors", "gemma4_assistant", "text-generation", "any-to-any", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_assistant", "total_params_b": 0.469518596, "active_params_b": 0.469518596, "layers": 4, "attention_heads": 32, "kv_heads": 16, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 0.939037192, "active_weight_gb": 0.939037192, "kv_alloc_gb_per_1k_tokens": 0.053248, "kv_read_gb_per_1k_tokens": 0.053248, "notes": "Manually reviewed on 2026-07-05 from current Hugging Face API metadata, served config, generation config, model card, and direct safetensors header range read. This repo is a Gemma 4 MTP speculative drafter used as assistant_model with google/gemma-4-31B-it. The profile records the exact 0.939037192 GB resident BF16 drafter artifact and intentionally disables production throughput because Bounds Engine v1 lacks paired target+assistant speculative decoding." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-4-31B-it-assistant" }, { "label": "Model config", "url": "https://huggingface.co/google/gemma-4-31B-it-assistant/raw/34ef9f029d1c52bccac2def222523af32f3ccd0f/config.json" }, { "label": "Generation config", "url": "https://huggingface.co/google/gemma-4-31B-it-assistant/raw/34ef9f029d1c52bccac2def222523af32f3ccd0f/generation_config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-31B-it-assistant" }, { "label": "Target model card", "url": "https://huggingface.co/google/gemma-4-31B-it" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/google/gemma-4-31B-it-assistant/resolve/34ef9f029d1c52bccac2def222523af32f3ccd0f/model.safetensors" } ] }, { "id": "unsloth/gemma-4-31B-it-GGUF", "name": "Gemma 4 31B IT GGUF IQ4_NL", "short_name": "Gemma 4 31B IT GGUF", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 607972, "tags": [ "gguf", "gemma4", "unsloth", "gemma", "google", "image-text-to-text", "base_model:google/gemma-4-31B-it", "base_model:quantized:google/gemma-4-31B-it", "license:apache-2.0", "endpoints_compatible", "region:us", "imatrix", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 30.697345596, "active_params_b": 30.697345596, "layers": 60, "attention_heads": 32, "kv_heads": 16, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF IQ4_NL", "weight_bytes_per_param": 0.56316493821709, "resident_weight_gb": 17.287668736, "active_weight_gb": 17.271836544, "kv_alloc_gb_per_1k_tokens": 0.028868608, "kv_read_gb_per_1k_tokens": 0.0466944, "notes": "Manually audited on 2026-07-05 from current HF API metadata, pinned config, model card, Google base config comparison, linked-object sizes, and a direct GGUF header range read. The selected artifact is gemma-4-31B-it-IQ4_NL.gguf because it matches API gguf.totalFileSize. Bound weights use exact selected file bytes for residency and exact tensor spans for swept text-decode traffic; layered Gemma 4 KV supersedes the generated full-context estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/gemma-4-31B-it-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/gemma-4-31B-it-GGUF" }, { "label": "Pinned model config", "url": "https://huggingface.co/unsloth/gemma-4-31B-it-GGUF/raw/8906b3db2e669a0b1d6293c315d3f9fbf934a86d/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-31B-it" }, { "label": "Selected IQ4_NL GGUF tensor-index audit", "url": "https://huggingface.co/unsloth/gemma-4-31B-it-GGUF/resolve/8906b3db2e669a0b1d6293c315d3f9fbf934a86d/gemma-4-31B-it-IQ4_NL.gguf" } ] }, { "id": "lmstudio-community/Qwen3.6-35B-A3B-MLX-8bit", "name": "Qwen3.6-35B-A3B-MLX-8bit 8-bit", "short_name": "Qwen3.6 35B A3B MLX 8bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 629948, "tags": [ "transformers", "safetensors", "qwen3_5_moe", "image-text-to-text", "mlx", "conversational", "base_model:Qwen/Qwen3.6-35B-A3B", "base_model:quantized:Qwen/Qwen3.6-35B-A3B", "license:apache-2.0", "endpoints_compatible", "8-bit", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_5_moe_text", "total_params_b": 35.9518, "active_params_b": 3, "layers": 40, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "MLX 8-bit", "weight_bytes_per_param": 1.0492139106247809, "resident_weight_gb": 37.721128672, "active_weight_gb": 3.131668736, "kv_alloc_gb_per_1k_tokens": 0.0649216, "kv_read_gb_per_1k_tokens": 0.06496256, "notes": "Manually audited on 2026-07-05 from current HF API metadata, pinned served MLX config, base Qwen config comparison, safetensors index, direct shard header reads, and Transformers qwen3_5 runtime review. Header bytes record 37.721128672 GB resident, 36.287641856 GB ordinary text resident, 34.22552064 GB packed routed expert tensors, 2.062121216 GB fixed text traffic, and 1.433486816 GB resident-only embedding plus vision tensors. Bounds use the profiled MoE adapter and DeltaNet recurrent-state KV adapter, not this generated summary row." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-35B-A3B-MLX-8bit" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3.6-35B-A3B-MLX-8bit" }, { "label": "Pinned model config", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-35B-A3B-MLX-8bit/raw/ea4ab1544e989b4fa38711453e0296cd7b03ac43/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B" }, { "label": "Safetensors shard header audit", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-35B-A3B-MLX-8bit/raw/ea4ab1544e989b4fa38711453e0296cd7b03ac43/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen2-0.5B-Instruct", "name": "Qwen2 0.5B Instruct BF16", "short_name": "Qwen2 0.5B Instruct", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 555044, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "chat", "conversational", "en", "base_model:Qwen/Qwen2-0.5B", "base_model:finetune:Qwen/Qwen2-0.5B", "license:apache-2.0", "eval-results", "text-generation-inference", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 0.494032768, "active_params_b": 0.494032768, "layers": 24, "attention_heads": 14, "kv_heads": 2, "head_dim": 64, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 0.988065536, "active_weight_gb": 0.988065536, "kv_alloc_gb_per_1k_tokens": 0.012288, "kv_read_gb_per_1k_tokens": 0.012288, "notes": "Manually audited on 2026-07-05 from current HF API metadata, served config, model card, base config comparison, and direct safetensors header range read. The Instruct repo preserves the base tensor geometry but records 32768 max position embeddings instead of the base repo's 131072. Header bytes record 494032768 BF16 parameters / 0.988065536 GB resident, including the tied model.embed_tokens.weight output projection." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2-0.5B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2-0.5B-Instruct" }, { "label": "Pinned model config", "url": "https://huggingface.co/Qwen/Qwen2-0.5B-Instruct/raw/c540970f9e29518b1d8f06ab8b24cba66ad77b6d/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2-0.5B" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/Qwen/Qwen2-0.5B-Instruct/resolve/c540970f9e29518b1d8f06ab8b24cba66ad77b6d/model.safetensors" } ] }, { "id": "microsoft/phi-2", "name": "phi-2 F16", "short_name": "phi 2", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 787841, "tags": [ "transformers", "safetensors", "phi", "text-generation", "nlp", "code", "en", "license:mit", "text-generation-inference", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "phi (PhiForCausalLM)", "total_params_b": 2.77968384, "active_params_b": 2.64861184, "layers": 32, "attention_heads": 32, "kv_heads": 32, "head_dim": 80, "max_context_tokens": 2048, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "F16", "weight_bytes_per_param": 2, "resident_weight_gb": 5.55936768, "active_weight_gb": 5.29722368, "kv_alloc_gb_per_1k_tokens": 0.32768, "kv_read_gb_per_1k_tokens": 0.32768, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, model card, pinned config, official Transformers 4.37.0 Phi implementation, safetensors index, and direct range-read shard headers. Bound weights separate the 5.559367680 GB resident F16 checkpoint from the 5.297223680 GB ordinary text-decode swept subset; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/microsoft/phi-2" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/microsoft/phi-2" }, { "label": "Pinned model config", "url": "https://huggingface.co/microsoft/phi-2/raw/810d367871c1d460086d9f82db8696f2e0a0fcd0/config.json" }, { "label": "Transformers Phi implementation", "url": "https://raw.githubusercontent.com/huggingface/transformers/v4.37.0/src/transformers/models/phi/modeling_phi.py" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/microsoft/phi-2/resolve/810d367871c1d460086d9f82db8696f2e0a0fcd0/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen3-VL-30B-A3B-Instruct", "name": "Qwen3-VL-30B-A3B-Instruct BF16", "short_name": "Qwen3 VL 30B A3B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 566869, "tags": [ "transformers", "safetensors", "qwen3_vl_moe", "image-text-to-text", "conversational", "arxiv:2505.09388", "arxiv:2502.13923", "arxiv:2409.12191", "arxiv:2308.12966", "license:apache-2.0", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_vl_moe_text", "total_params_b": 31.070754032, "active_params_b": 3.041867776, "layers": 48, "attention_heads": 32, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 62.141508064, "active_weight_gb": 6.083735552, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, model card, pinned config, official Transformers 4.57.0 Qwen3 VL MoE implementation, safetensors index, and direct range-read shard headers. Bound weights separate the 62.141508064 GB resident BF16 checkpoint from the 6.083735552 GB single-token ordinary text-decode traffic; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-VL-30B-A3B-Instruct" }, { "label": "Pinned model config", "url": "https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct/raw/9c4b90e1e4ba969fd3b5378b57d966d725f1b86c/config.json" }, { "label": "Transformers 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" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct/resolve/9c4b90e1e4ba969fd3b5378b57d966d725f1b86c/model.safetensors.index.json" } ] }, { "id": "hugging-quants/Llama-3.2-1B-Instruct-Q8_0-GGUF", "name": "Llama-3.2-1B-Instruct-Q8_0-GGUF GGUF 8-bit", "short_name": "Llama 3.2 1B Instruct Q8 0", "license": "llama3.2", "hf_pipeline_tag": "text-generation", "hf_downloads": 573463, "tags": [ "gguf", "facebook", "meta", "pytorch", "llama", "llama-3", "llama-cpp", "gguf-my-repo", "text-generation", "en", "de", "fr", "it", "pt", "hi", "es", "th", "base_model:meta-llama/Llama-3.2-1B-Instruct", "base_model:quantized:meta-llama/Llama-3.2-1B-Instruct", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "llama GGUF", "total_params_b": 1.235814432, "active_params_b": 1.235814432, "layers": 16, "attention_heads": 32, "kv_heads": 8, "head_dim": 64, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF 8-bit", "weight_bytes_per_param": 1.0689947987271928, "resident_weight_gb": 1.3210792, "active_weight_gb": 1.313251456, "kv_alloc_gb_per_1k_tokens": 0.032768, "kv_read_gb_per_1k_tokens": 0.032768, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, gated base-model API/config access checks, linked-object HEAD check, and direct GGUF header range read. Bound weights use the exact selected llama-3.2-1b-instruct-q8_0.gguf file size for residency and exact tensor spans for ordinary text decode; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/hugging-quants/Llama-3.2-1B-Instruct-Q8_0-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/hugging-quants/Llama-3.2-1B-Instruct-Q8_0-GGUF" }, { "label": "Base model API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.2-1B-Instruct" }, { "label": "Gated base config access check", "url": "https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct/raw/main/config.json" }, { "label": "Selected GGUF tensor index audit", "url": "https://huggingface.co/hugging-quants/Llama-3.2-1B-Instruct-Q8_0-GGUF/resolve/8790ecfe0c6c7a20f2a532acdfc1b5d98e1bfa8b/llama-3.2-1b-instruct-q8_0.gguf" } ] }, { "id": "Qwen/Qwen3-8B-FP8", "name": "Qwen3-8B-FP8 FP8", "short_name": "Qwen3 8B FP8", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 609936, "tags": [ "transformers", "safetensors", "qwen3", "text-generation", "conversational", "arxiv:2309.00071", "arxiv:2505.09388", "base_model:Qwen/Qwen3-8B", "base_model:quantized:Qwen/Qwen3-8B", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "fp8", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 8.191159296, "active_params_b": 7.56882944, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "FP8", "weight_bytes_per_param": 1.1520409782053884, "resident_weight_gb": 9.436551168, "active_weight_gb": 8.191891456, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, pinned served FP8 config, audited BF16 base config/profile comparison, Transformers Qwen3 implementation review, safetensors index, and direct shard header range reads. Bound weights use exact direct tensor payload bytes: 9.436551168 GB resident and 8.191891456 GB ordinary text-decode swept traffic after excluding only model.embed_tokens.weight. The audited profile supersedes the generated rounded FP8 estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-8B-FP8" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-8B-FP8" }, { "label": "Pinned model config", "url": "https://huggingface.co/Qwen/Qwen3-8B-FP8/raw/220b46e3b2180893580a4454f21f22d3ebb187d3/config.json" }, { "label": "Pinned base model config", "url": "https://huggingface.co/Qwen/Qwen3-8B/raw/b968826d9c46dd6066d109eabc6255188de91218/config.json" }, { "label": "Transformers Qwen3 implementation", "url": "https://raw.githubusercontent.com/huggingface/transformers/v4.51.0/src/transformers/models/qwen3/modeling_qwen3.py" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/Qwen/Qwen3-8B-FP8/resolve/220b46e3b2180893580a4454f21f22d3ebb187d3/model.safetensors.index.json" } ] }, { "id": "unsloth/Qwen3-30B-A3B-Instruct-2507-GGUF", "name": "Qwen3-30B-A3B-Instruct-2507-GGUF GGUF", "short_name": "Qwen3 30B A3B Instruct 2507", "license": "apache-2.0", "hf_pipeline_tag": "unknown", "hf_downloads": 468009, "tags": [ "transformers", "gguf", "qwen", "qwen3", "unsloth", "arxiv:2505.09388", "base_model:Qwen/Qwen3-30B-A3B-Instruct-2507", "base_model:quantized:Qwen/Qwen3-30B-A3B-Instruct-2507", "license:apache-2.0", "endpoints_compatible", "region:us", "imatrix", "conversational" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_moe (Qwen3MoeForCausalLM)", "total_params_b": 30.532122624, "active_params_b": 3.041867776, "layers": 48, "attention_heads": 32, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "GGUF IQ4_NL", "weight_bytes_per_param": 0.5669694855212173, "resident_weight_gb": 17.310781856, "active_weight_gb": 1.841543168, "expert_param_b": 0.226492416, "fixed_param_b": 1.229928448, "shared_expert_param_b": 0, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, model card, audited BF16 base profile/config, selected linked-object metadata, and a direct GGUF header range read. The API-selected artifact is Qwen3-30B-A3B-Instruct-2507-IQ4_NL.gguf at 17.310781856 GB resident. Ordinary single-session MoE text-decode traffic is 1.841543168 GB: 0.822327296 GB fixed traffic plus 8 routed expert groups at 0.127401984 GB each. The audited profile supersedes the generated BF16-derived estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/Qwen3-30B-A3B-Instruct-2507-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3-30B-A3B-Instruct-2507-GGUF" }, { "label": "Audited base model config", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507/raw/0d7cf23991f47feeb3a57ecb4c9cee8ea4a17bfe/config.json" }, { "label": "Selected IQ4_NL GGUF tensor-index audit", "url": "https://huggingface.co/unsloth/Qwen3-30B-A3B-Instruct-2507-GGUF/resolve/eea7b2be5805a5f151f8847ede8e5f9a9284bf77/Qwen3-30B-A3B-Instruct-2507-IQ4_NL.gguf" } ] }, { "id": "zai-org/GLM-4.1V-9B-Thinking", "name": "GLM-4.1V-9B-Thinking BF16", "short_name": "GLM 4.1V 9B Thinking", "license": "mit", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 580375, "tags": [ "transformers", "safetensors", "glm4v", "image-text-to-text", "reasoning", "conversational", "en", "zh", "arxiv:2507.01006", "base_model:zai-org/GLM-4-9B-0414", "base_model:finetune:zai-org/GLM-4-9B-0414", "license:mit", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "glm4v_text", "total_params_b": 10.292777472, "active_params_b": 8.779522048, "layers": 40, "attention_heads": 32, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 65536, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 20.585554944, "active_weight_gb": 17.559044096, "kv_alloc_gb_per_1k_tokens": 0.04096, "kv_read_gb_per_1k_tokens": 0.04096, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, pinned served GLM4V config, GLM-4-9B-0414 base config comparison, Transformers GLM4V implementation review, safetensors index, and direct shard header range reads. Bound weights separate the exact 20.585554944 GB resident BF16 multimodal package from 17.559044096 GB ordinary text-decode swept traffic; model.visual.* and model.language_model.embed_tokens.weight are resident-only for ordinary text-token decode. The audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/zai-org/GLM-4.1V-9B-Thinking" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/zai-org/GLM-4.1V-9B-Thinking" }, { "label": "Pinned model config", "url": "https://huggingface.co/zai-org/GLM-4.1V-9B-Thinking/raw/9e9a4c5e94f4a095c353f4152d520a2644a553b2/config.json" }, { "label": "Pinned base model config", "url": "https://huggingface.co/zai-org/GLM-4-9B-0414/raw/645b8482494e31b6b752272bf7f7f273ef0f3caf/config.json" }, { "label": "Transformers GLM4V implementation", "url": "https://raw.githubusercontent.com/huggingface/transformers/v4.57.1/src/transformers/models/glm4v/modeling_glm4v.py" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/zai-org/GLM-4.1V-9B-Thinking/resolve/9e9a4c5e94f4a095c353f4152d520a2644a553b2/model.safetensors.index.json" } ] }, { "id": "lmstudio-community/Qwen3.6-35B-A3B-MLX-6bit", "name": "Qwen3.6-35B-A3B-MLX-6bit BF16/U32", "short_name": "Qwen3.6 35B A3B MLX 6bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 618161, "tags": [ "transformers", "safetensors", "qwen3_5_moe", "image-text-to-text", "mlx", "conversational", "base_model:Qwen/Qwen3.6-35B-A3B", "base_model:quantized:Qwen/Qwen3.6-35B-A3B", "license:apache-2.0", "endpoints_compatible", "6-bit", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_5_moe_text", "total_params_b": 35.9518, "active_params_b": 3, "layers": 40, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "MLX 6-bit affine", "weight_bytes_per_param": 0.8083469899142742, "resident_weight_gb": 29.061529312, "active_weight_gb": 4.350966016, "kv_alloc_gb_per_1k_tokens": 0.02048, "kv_read_gb_per_1k_tokens": 0.02048, "kv_alloc_fixed_gb": 0.06488064, "kv_read_fixed_gb": 0.06488064, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, pinned served MLX config, Qwen3.6 35B A3B base config comparison, model card, safetensors index, direct shard header range reads, and the Transformers qwen3_5 implementation. Bound weights separate the exact 29.061529312 GB resident MLX package from the 4.350966016 GB single-session active ordinary text-decode traffic; the audited profile supersedes the generated BF16/U32 estimate and adds the fixed DeltaNet runtime state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-35B-A3B-MLX-6bit" }, { "label": "Model config", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-35B-A3B-MLX-6bit/raw/ee1b3d2dc76e36e305a48f24b168a903fec6bc0d/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3.6-35B-A3B-MLX-6bit" }, { "label": "Pinned base config", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/995ad96eacd98c81ed38be0c5b274b04031597b0/config.json" }, { "label": "Safetensors index", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-35B-A3B-MLX-6bit/raw/ee1b3d2dc76e36e305a48f24b168a903fec6bc0d/model.safetensors.index.json" } ] }, { "id": "microsoft/Phi-3-mini-4k-instruct", "name": "Phi-3-mini-4k-instruct BF16", "short_name": "Phi 3 mini 4k", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 578189, "tags": [ "transformers", "safetensors", "phi3", "text-generation", "nlp", "code", "conversational", "custom_code", "en", "fr", "license:mit", "eval-results", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "phi3 (Phi3ForCausalLM)", "total_params_b": 3.821079552, "active_params_b": 3.821079552, "layers": 32, "attention_heads": 32, "kv_heads": 32, "head_dim": 96, "max_context_tokens": 4096, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 7.642159104, "active_weight_gb": 7.445157888, "kv_alloc_gb_per_1k_tokens": 0.393216, "kv_read_gb_per_1k_tokens": 0.393216, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, model card, served config, remote modeling code, safetensors index, and direct safetensors header range reads. Bound weights separate the 7.642159104 GB resident BF16 checkpoint from the 7.445157888 GB ordinary text-decode swept subset. KV display fields show first-1K text-cache traffic; the audited profile supersedes these legacy adapter fields for production bounds and caps allocation/read traffic at the 2047-token sliding window." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct" }, { "label": "Model config", "url": "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/raw/main/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/microsoft/Phi-3-mini-4k-instruct" }, { "label": "Pinned config", "url": "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/raw/f39ac1d28e925b323eae81227eaba4464caced4e/config.json" }, { "label": "Remote modeling code", "url": "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/raw/f39ac1d28e925b323eae81227eaba4464caced4e/modeling_phi3.py" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/raw/f39ac1d28e925b323eae81227eaba4464caced4e/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen3-4B-Thinking-2507", "name": "Qwen3-4B-Thinking-2507 BF16", "short_name": "Qwen3 4B Thinking 2507", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 554589, "tags": [ "transformers", "safetensors", "qwen3", "text-generation", "conversational", "arxiv:2505.09388", "license:apache-2.0", "eval-results", "text-generation-inference", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 4.022468096, "active_params_b": 4.022468096, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 8.044936192, "active_weight_gb": 8.044936192, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, model card, served config, sibling config comparison, safetensors index, and direct safetensors header range reads. The repo stores 4.022468096B BF16 parameters / 8.044936192 GB across 398 tensors. The tied model.embed_tokens.weight is also the output projection, so the dense v1 profile charges all resident weights as text-decode traffic." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-4B-Thinking-2507" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-4B-Thinking-2507/raw/main/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-4B-Thinking-2507" }, { "label": "Pinned config", "url": "https://huggingface.co/Qwen/Qwen3-4B-Thinking-2507/raw/768f209d9ea81521153ed38c47d515654e938aea/config.json" }, { "label": "Sibling config comparison", "url": "https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Qwen/Qwen3-4B-Thinking-2507/raw/768f209d9ea81521153ed38c47d515654e938aea/model.safetensors.index.json" } ] }, { "id": "unsloth/Qwen3.6-27B-GGUF", "name": "Qwen3.6-27B-GGUF GGUF", "short_name": "Qwen3.6 27B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 588160, "tags": [ "transformers", "gguf", "unsloth", "qwen", "qwen3_5", "image-text-to-text", "base_model:Qwen/Qwen3.6-27B", "base_model:quantized:Qwen/Qwen3.6-27B", "license:apache-2.0", "endpoints_compatible", "region:us", "imatrix", "conversational" ], "architecture": { "type": "Dense multimodal transformer", "detail": "qwen3_5_text (API-selected IQ4_NL GGUF)", "total_params_b": 26.895998464, "active_params_b": 25.624600064, "layers": 64, "attention_heads": 24, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF IQ4_NL", "weight_bytes_per_param": 0.5975525469155529, "resident_weight_gb": 16.071772384, "active_weight_gb": 15.345616896, "kv_alloc_gb_per_1k_tokens": 0.065536, "kv_read_gb_per_1k_tokens": 0.065536, "kv_alloc_fixed_gb": 0.154927104, "kv_read_fixed_gb": 0.154927104, "notes": "Manually audited on 2026-07-05 from HF API commit 82d411acf4a06cfb8d9b073a5211bf410bfc29bf, model card, pinned Qwen base config, linked GGUF file sizes, and a GGUF tensor-index range read. The authoritative bounds profile uses the API-selected Qwen3.6-27B-IQ4_NL.gguf artifact: 16.071772384 GB resident, 15.345616896 GB swept ordinary text traffic, 0.726155488 GB resident-only token embedding plus GGUF metadata/header/file overhead, and Qwen3.6 hybrid KV plus fixed DeltaNet state. Separate mmproj sidecars are not included unless explicitly loaded." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/Qwen3.6-27B-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.6-27B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3.6-27B-GGUF" }, { "label": "Pinned base config", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/unsloth/Qwen3.6-27B-GGUF/resolve/82d411acf4a06cfb8d9b073a5211bf410bfc29bf/Qwen3.6-27B-IQ4_NL.gguf" } ] }, { "id": "google/gemma-4-26B-A4B", "name": "gemma-4-26B-A4B BF16", "short_name": "gemma 4 26B A4B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 720641, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "gemma4_text", "total_params_b": 26.544131376, "active_params_b": 4, "layers": 30, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 51.611872412, "active_weight_gb": 7.64506118, "kv_alloc_gb_per_1k_tokens": 0.20992, "kv_read_gb_per_1k_tokens": 0.20992, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, pinned config, model card, safetensors index, and direct shard header range reads. The authoritative bounds profile records 51.611872412 GB resident BF16 tensors, 4.79031302 GB fixed text traffic, 0.35684352 GB per routed expert, 8 routed experts per token, 128 routed experts, 1 shared expert charged in fixed traffic, and layered Gemma 4 hybrid KV. Legacy per-1k KV fields show the short-context unsaturated coefficient; long-context calculation must use the profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-4-26B-A4B" }, { "label": "Model config", "url": "https://huggingface.co/google/gemma-4-26B-A4B/raw/f1102d7de421741c6eafcda46d1806a7a65b83a3/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-26B-A4B" }, { "label": "Safetensors index", "url": "https://huggingface.co/google/gemma-4-26B-A4B/raw/f1102d7de421741c6eafcda46d1806a7a65b83a3/model.safetensors.index.json" }, { "label": "Instruction-tuned sibling config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it/raw/20da991ab4afab98e8f910c4a2e8f4fbefc404ad/config.json" } ] }, { "id": "yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF", "name": "gemma-4-12B-coder-fable5-composer2.5-v1-GGUF GGUF", "short_name": "gemma 4 12B coder fable5 composer2.5 v1", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 641260, "tags": [ "gguf", "gemma4", "coding", "code", "reasoning", "thinking", "llama.cpp", "local-llm", "text-generation", "base_model:google/gemma-4-12B-it", "base_model:quantized:google/gemma-4-12B-it", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_unified_text", "total_params_b": 11.907350576, "active_params_b": 11.907350576, "layers": 48, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q4_K_M", "weight_bytes_per_param": 0.6199012632480672, "resident_weight_gb": 7.381381664, "active_weight_gb": 7.365559808, "kv_alloc_gb_per_1k_tokens": 0.335872, "kv_read_gb_per_1k_tokens": 0.335872, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, model card, linked full-precision fine-tune config, Google base config comparison, linked GGUF file sizes, and a direct GGUF tensor-index range read. The authoritative bounds profile uses the card-recommended gemma4-coding-Q4_K_M.gguf artifact: 7.381381664 GB resident, 7.365559808 GB swept text traffic, 0.015821856 GB resident-only GGUF metadata/header bytes, and layered Gemma 4 hybrid KV. API gguf.totalFileSize selects the smaller Q2_K file, but the model card calls Q4_K_M the sweet spot and uses it in the example command. Legacy per-1k KV fields show the short-context unsaturated coefficient; long-context calculation must use the profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF" }, { "label": "Full-precision fine-tune config", "url": "https://huggingface.co/yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1/raw/237e5d8e72fb70e2c4d5df2676bc95308aad279c/config.json" }, { "label": "Google base config", "url": "https://huggingface.co/google/gemma-4-12B-it/raw/5926caa4ec0cac5cbfadaf4077420520de1d5205/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF/resolve/1380be1796e559fca96b4107599285cab3ddbb92/gemma4-coding-Q4_K_M.gguf" } ] }, { "id": "microsoft/Phi-4-mini-instruct", "name": "Phi-4-mini-instruct BF16", "short_name": "Phi 4 mini", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 513426, "tags": [ "transformers", "safetensors", "phi3", "text-generation", "nlp", "code", "conversational", "custom_code", "multilingual", "ar", "zh", "cs", "da", "nl", "en", "fi", "fr", "de", "he", "hu", "it", "ja", "ko", "no", "pl", "pt", "ru", "es", "sv", "th", "tr", "uk", "arxiv:2503.01743", "license:mit", "eval-results", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "phi3 (Phi3ForCausalLM)", "total_params_b": 3.83602176, "active_params_b": 3.83602176, "layers": 32, "attention_heads": 24, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 7.67204352, "active_weight_gb": 7.67204352, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, config.json, model card, remote modeling code, safetensors index, and direct shard header range reads. The authoritative bounds profile records 7.672043520 GB resident and swept BF16 tensors. The repo ties input and output embeddings and stores no separate lm_head.weight, so model.embed_tokens.weight is charged as swept output-projection traffic for ordinary text decode. The config's 262144-token sliding window is larger than the supported 131072-token context, so KV traffic is full-context within the valid context." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/microsoft/Phi-4-mini-instruct" }, { "label": "Model config", "url": "https://huggingface.co/microsoft/Phi-4-mini-instruct/raw/cfbefacb99257ffa30c83adab238a50856ac3083/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/microsoft/Phi-4-mini-instruct" }, { "label": "Remote modeling code", "url": "https://huggingface.co/microsoft/Phi-4-mini-instruct/raw/cfbefacb99257ffa30c83adab238a50856ac3083/modeling_phi3.py" }, { "label": "Safetensors index", "url": "https://huggingface.co/microsoft/Phi-4-mini-instruct/resolve/cfbefacb99257ffa30c83adab238a50856ac3083/model.safetensors.index.json" } ] }, { "id": "MiniMaxAI/MiniMax-M3-MXFP8", "name": "MiniMax-M3-MXFP8 FP8", "short_name": "MiniMax M3 MXFP8", "license": "other", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 699967, "tags": [ "transformers", "safetensors", "minimax_m3_vl", "image-text-to-text", "multimodal", "moe", "agent", "coding", "video", "conversational", "custom_code", "arxiv:2606.13392", "base_model:MiniMaxAI/MiniMax-M3", "base_model:quantized:MiniMaxAI/MiniMax-M3", "license:other", "endpoints_compatible", "mxfp8", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "minimax_m3_vl (MiniMaxM3SparseForCausalLM)", "total_params_b": 440.27984576, "active_params_b": 23, "layers": 60, "attention_heads": 64, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 1048576, "routed_experts": 128, "routed_experts_per_token": 4, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "MXFP8", "weight_bytes_per_param": 1.0078637243865287, "resident_weight_gb": 443.74208512, "active_weight_gb": 26.830826496, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manual correction on 2026-07-05 from live Hugging Face API metadata, served MXFP8 config, BF16 base config comparison, model card MSA notes, safetensors index, and direct shard-header range reads. The model card says about 23B activated parameters; active_weight_gb records exact single-token ordinary text weight traffic from headers: 13.517319168 GB fixed traffic plus 4 routed expert groups at 3.328376832 GB each. KV fields are zeroed in this legacy row because MiniMax Sparse Attention allocation/read traffic is not audited; the authoritative profile is fail-closed and returns unsupported_profile for production bounds." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MiniMaxAI/MiniMax-M3-MXFP8" }, { "label": "Model card", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M3-MXFP8" }, { "label": "Pinned served config", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M3-MXFP8/raw/ca165902e868fd015fab5acaff776643be00dc6e/config.json" }, { "label": "Pinned base config", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M3/raw/fa6849585c885749b76c0b1f299841b231e4cbfa/config.json" }, { "label": "Safetensors index", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M3-MXFP8/resolve/ca165902e868fd015fab5acaff776643be00dc6e/model.safetensors.index.json" } ] }, { "id": "QuantTrio/Qwen3.6-35B-A3B-AWQ", "name": "Qwen3.6-35B-A3B-AWQ 4-bit", "short_name": "Qwen3.6 35B A3B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 695633, "tags": [ "transformers", "safetensors", "qwen3_5_moe", "image-text-to-text", "vLLM", "AWQ", "conversational", "base_model:Qwen/Qwen3.6-35B-A3B", "base_model:quantized:Qwen/Qwen3.6-35B-A3B", "license:apache-2.0", "endpoints_compatible", "4-bit", "awq", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_5_moe_text", "total_params_b": 35.951822704, "active_params_b": 3, "layers": 40, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "4-bit", "weight_bytes_per_param": 0.5, "resident_weight_gb": 25.44858288, "active_weight_gb": 4.481770752, "kv_alloc_gb_per_1k_tokens": 0.08536064, "kv_read_gb_per_1k_tokens": 0.08536064, "notes": "Manual correction on 2026-07-05 from live Hugging Face API metadata, model card, pinned AWQ config, base config comparison, safetensors index, and direct range-read shard headers. Bound weights separate the 25.448582880 GB resident AWQ package from 4.481770752 GB single-token ordinary text traffic: 3.921536256 GB fixed traffic plus 8 routed expert groups at 0.070029312 GB each. KV display fields show first-1K text-cache plus fixed DeltaNet state traffic; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/QuantTrio/Qwen3.6-35B-A3B-AWQ" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/QuantTrio/Qwen3.6-35B-A3B-AWQ" }, { "label": "Pinned AWQ config", "url": "https://huggingface.co/QuantTrio/Qwen3.6-35B-A3B-AWQ/raw/119886a1072372348f73ef0df2d801cdcc0f455b/config.json" }, { "label": "Base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/995ad96eacd98c81ed38be0c5b274b04031597b0/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/QuantTrio/Qwen3.6-35B-A3B-AWQ/resolve/119886a1072372348f73ef0df2d801cdcc0f455b/model.safetensors.index.json" } ] }, { "id": "cyankiwi/Qwen3.5-27B-AWQ-4bit", "name": "Qwen3.5-27B-AWQ-4bit 4-bit", "short_name": "Qwen3.5 27B AWQ 4bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 557518, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "conversational", "base_model:Qwen/Qwen3.5-27B", "base_model:quantized:Qwen/Qwen3.5-27B", "license:apache-2.0", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 27.781427952, "active_params_b": 25.624600064, "layers": 64, "attention_heads": 24, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "compressed-tensors AWQ INT4", "weight_bytes_per_param": 0.7219841762869511, "resident_weight_gb": 20.057751376, "active_weight_gb": 16.279196928, "kv_alloc_gb_per_1k_tokens": 0.065536, "kv_read_gb_per_1k_tokens": 0.065536, "kv_alloc_fixed_gb": 0.154927104, "kv_read_fixed_gb": 0.154927104, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, pinned served compressed-tensors config, Qwen3.5 27B base config comparison, recipe, safetensors index, direct shard header range reads, and the Transformers qwen3_5 implementation. Bound weights separate the exact 20.057751376 GB resident package from 16.279196928 GB ordinary text-decode swept traffic; the audited profile supersedes the generated ideal 4-bit estimate and adds the fixed DeltaNet runtime state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/cyankiwi/Qwen3.5-27B-AWQ-4bit" }, { "label": "Model config", "url": "https://huggingface.co/cyankiwi/Qwen3.5-27B-AWQ-4bit/raw/65497df9ca706cd6aa9cc8374fc259f39e78672d/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-27B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3.5-27B-AWQ-4bit" }, { "label": "Pinned base config", "url": "https://huggingface.co/Qwen/Qwen3.5-27B/raw/fc05daec18b0a78c049392ed2e771dde82bdf654/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/cyankiwi/Qwen3.5-27B-AWQ-4bit/resolve/65497df9ca706cd6aa9cc8374fc259f39e78672d/model.safetensors.index.json" } ] }, { "id": "deepseek-ai/deepseek-coder-7b-instruct-v1.5", "name": "deepseek-coder-7b-instruct-v1.5 BF16", "short_name": "deepseek coder 7b instruct v1.5", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 549396, "tags": [ "transformers", "safetensors", "llama", "text-generation", "conversational", "license:other", "text-generation-inference", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "llama (LlamaForCausalLM)", "total_params_b": 6.910365696, "active_params_b": 6.490935296, "layers": 30, "attention_heads": 32, "kv_heads": 32, "head_dim": 128, "max_context_tokens": 4096, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 13.820731392, "active_weight_gb": 12.981870592, "kv_alloc_gb_per_1k_tokens": 0.49152, "kv_read_gb_per_1k_tokens": 0.49152, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, model card, pinned served config, safetensors index, and direct shard header range reads. Bound weights separate the exact 13.820731392 GB resident BF16 package from 12.981870592 GB ordinary text-decode swept traffic; the audited profile supersedes the generated rounded dense estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/deepseek-ai/deepseek-coder-7b-instruct-v1.5" }, { "label": "Model config", "url": "https://huggingface.co/deepseek-ai/deepseek-coder-7b-instruct-v1.5/raw/2a050a4c59d687a85324d32e147517992117ed30/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/deepseek-coder-7b-instruct-v1.5" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/deepseek-ai/deepseek-coder-7b-instruct-v1.5/resolve/2a050a4c59d687a85324d32e147517992117ed30/model.safetensors.index.json" } ] }, { "id": "speakleash/Bielik-11B-v3.0-Instruct-awq", "name": "Bielik-11B-v3.0-Instruct-awq AWQ 4-bit", "short_name": "Bielik 11B v3.0 Instruct", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 628428, "tags": [ "transformers", "safetensors", "llama", "text-generation", "finetuned", "gguf", "conversational", "multilingual", "pl", "en", "sq", "bel", "bs", "bg", "hr", "cs", "da", "et", "fi", "fr", "el", "es", "is", "lt", "nl", "de", "no", "pt", "ru", "ro", "sr", "hbs", "sv", "sk", "sl", "tr", "uk", "hu", "it", "lv", "base_model:speakleash/Bielik-11B-v3.0-Instruct", "base_model:quantized:speakleash/Bielik-11B-v3.0-Instruct", "license:apache-2.0", "text-generation-inference", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "llama (LlamaForCausalLM)", "total_params_b": 11.168796672, "active_params_b": 11.037200384, "layers": 50, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "compressed-tensors AWQ INT4", "weight_bytes_per_param": 0.5544738189679169, "resident_weight_gb": 6.192805344, "active_weight_gb": 5.929612768, "kv_alloc_gb_per_1k_tokens": 0.2048, "kv_read_gb_per_1k_tokens": 0.2048, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, model card, pinned served compressed-tensors config, base API metadata, safetensors index, and direct shard header range reads. Bound weights separate the exact 6.192805344 GB resident AWQ package from 5.929612768 GB ordinary text-decode swept traffic; the audited profile supersedes the generated GGUF label and ideal 4-bit estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/speakleash/Bielik-11B-v3.0-Instruct-awq" }, { "label": "Model config", "url": "https://huggingface.co/speakleash/Bielik-11B-v3.0-Instruct-awq/raw/0de6dd10ddcbd570b3844f1ae0e98f8e4800b548/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/speakleash/Bielik-11B-v3.0-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/speakleash/Bielik-11B-v3.0-Instruct-awq" }, { "label": "Base API metadata", "url": "https://huggingface.co/api/models/speakleash/Bielik-11B-v3.0-Instruct" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/speakleash/Bielik-11B-v3.0-Instruct-awq/resolve/0de6dd10ddcbd570b3844f1ae0e98f8e4800b548/model.safetensors.index.json" } ] }, { "id": "nvidia/Qwen3.5-397B-A17B-NVFP4", "name": "Qwen3.5-397B-A17B-NVFP4 NVFP4", "short_name": "Qwen3.5 397B A17B NVFP4", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 488162, "tags": [ "Model Optimizer", "safetensors", "qwen3_5_moe", "nvidia", "ModelOpt", "Qwen3.5", "quantized", "FP4", "fp4", "text-generation", "conversational", "base_model:Qwen/Qwen3.5-397B-A17B", "base_model:quantized:Qwen/Qwen3.5-397B-A17B", "license:apache-2.0", "modelopt", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_5_moe_text", "total_params_b": 403.3979, "active_params_b": 17, "layers": 60, "attention_heads": 32, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "NVFP4/FP8/BF16/F32", "weight_bytes_per_param": 0.5, "resident_weight_gb": 251.135201368, "active_weight_gb": 9.177853624, "kv_alloc_gb_per_1k_tokens": 0.01536, "kv_read_gb_per_1k_tokens": 0.01536, "kv_alloc_fixed_gb": 0.19316736, "kv_read_fixed_gb": 0.19316736, "notes": "Audited on 2026-07-05 from live Hugging Face metadata, pinned config/hf_quant_config, base config comparison, and direct safetensors header range reads; downloads=488162. Legacy catalog fields use exact stored resident bytes, exact single-sequence active ordinary text traffic, FP8 full-attention KV traffic, and fixed DeltaNet state. Production bounds use the self-contained profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/Qwen3.5-397B-A17B-NVFP4" }, { "label": "Model config", "url": "https://huggingface.co/nvidia/Qwen3.5-397B-A17B-NVFP4/raw/12b28061efd8145b11226e4fab2528ee4f39ef09/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-397B-A17B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/Qwen3.5-397B-A17B-NVFP4" }, { "label": "ModelOpt quantization config", "url": "https://huggingface.co/nvidia/Qwen3.5-397B-A17B-NVFP4/raw/12b28061efd8145b11226e4fab2528ee4f39ef09/hf_quant_config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/nvidia/Qwen3.5-397B-A17B-NVFP4/resolve/12b28061efd8145b11226e4fab2528ee4f39ef09/model.safetensors.index.json" } ] }, { "id": "google/gemma-4-12B-it-qat-q4_0-gguf", "name": "gemma-4-12B-it-qat-q4_0-gguf GGUF 4-bit", "short_name": "gemma 4 12B it qat q4 0", "license": "apache-2.0", "hf_pipeline_tag": "any-to-any", "hf_downloads": 564056, "tags": [ "transformers", "gguf", "any-to-any", "base_model:google/gemma-4-12B-it-qat-q4_0-unquantized", "base_model:quantized:google/gemma-4-12B-it-qat-q4_0-unquantized", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_unified_text", "total_params_b": 11.907350576, "active_params_b": 11.907350576, "layers": 48, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q4_0 QAT", "weight_bytes_per_param": 0.5858463378125703, "resident_weight_gb": 6.975877728, "active_weight_gb": 6.960055808, "kv_alloc_gb_per_1k_tokens": 0.335872, "kv_read_gb_per_1k_tokens": 0.335872, "notes": "Manually audited on 2026-07-05 from HF API commit f6e7774e6148da3b7f201e42ba37cf084c1db35f, Google QAT unquantized base config, model card, linked GGUF file sizes, and a GGUF tensor-index range read. The authoritative bounds profile uses the selected gemma-4-12b-it-qat-q4_0.gguf artifact: 6.975877728 GB resident, 6.960055808 GB swept text traffic, 0.01582192 GB resident-only GGUF metadata/header bytes, and layered Gemma 4 hybrid KV. The separate mmproj GGUF file is not included unless explicitly loaded. Legacy per-1k KV fields show the short-context unsaturated coefficient; long-context calculation must use the profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-4-12B-it-qat-q4_0-gguf" }, { "label": "Base model config", "url": "https://huggingface.co/google/gemma-4-12B-it-qat-q4_0-unquantized/raw/58540658b6c08edab2ddc1fbde7f28cc9987ced3/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-12B-it-qat-q4_0-unquantized" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-12B-it-qat-q4_0-gguf" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/google/gemma-4-12B-it-qat-q4_0-gguf/resolve/f6e7774e6148da3b7f201e42ba37cf084c1db35f/gemma-4-12b-it-qat-q4_0.gguf" } ] }, { "id": "google/gemma-4-31B", "name": "gemma-4-31B BF16", "short_name": "gemma 4 31B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 656084, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 31.273088876, "active_params_b": 30.69734534, "layers": 60, "attention_heads": 32, "kv_heads": 16, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 62.546177752, "active_weight_gb": 61.39469068, "kv_alloc_gb_per_1k_tokens": 0.83968, "kv_read_gb_per_1k_tokens": 0.83968, "notes": "Manually audited on 2026-07-05 from HF API commit 02e15e4990e8c452f8543fb26beff15b1daf8f3d, served config, instruction-tuned sibling config comparison, safetensors index, and direct shard header range reads. The authoritative bounds profile uses exact stored bytes: 62.546177752 GB resident, 61.39469068 GB swept ordinary text traffic, 1.151487072 GB resident-only vision/projection traffic, and layered Gemma 4 BF16 KV with separate sliding-layer K/V streams. Legacy per-1k KV fields show the short-context unsaturated coefficient; long-context calculation must use the profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-4-31B" }, { "label": "Model config", "url": "https://huggingface.co/google/gemma-4-31B/raw/02e15e4990e8c452f8543fb26beff15b1daf8f3d/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-31B" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/google/gemma-4-31B/resolve/02e15e4990e8c452f8543fb26beff15b1daf8f3d/model.safetensors.index.json" } ] }, { "id": "lmstudio-community/gemma-4-12B-it-GGUF", "name": "gemma-4-12B-it-GGUF GGUF", "short_name": "gemma 4 12B it", "license": "apache-2.0", "hf_pipeline_tag": "unknown", "hf_downloads": 409335, "tags": [ "gguf", "base_model:google/gemma-4-12B-it", "base_model:quantized:google/gemma-4-12B-it", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_unified_text", "total_params_b": 11.907350576, "active_params_b": 11.907350576, "layers": 48, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q4_K_M", "weight_bytes_per_param": 0.6199049820159312, "resident_weight_gb": 7.381384864, "active_weight_gb": 7.365559808, "kv_alloc_gb_per_1k_tokens": 0.335872, "kv_read_gb_per_1k_tokens": 0.335872, "notes": "Manually audited on 2026-07-05 from HF API commit ad4c92b806b5efd34055a9343e022736ad182634, Google Gemma 4 12B IT base config, model card, linked GGUF file sizes, and a GGUF tensor-index range read. The authoritative bounds profile uses the selected gemma-4-12B-it-Q4_K_M.gguf artifact: 7.381384864 GB resident, 7.365559808 GB swept text traffic, 0.015825056 GB resident-only GGUF metadata/header bytes, and layered Gemma 4 hybrid KV. The separate mmproj GGUF file is not included unless explicitly loaded. Legacy per-1k KV fields show the short-context unsaturated coefficient; long-context calculation must use the profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/gemma-4-12B-it-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/google/gemma-4-12B-it/raw/5926caa4ec0cac5cbfadaf4077420520de1d5205/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-12B-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-12B-it-GGUF" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/lmstudio-community/gemma-4-12B-it-GGUF/resolve/ad4c92b806b5efd34055a9343e022736ad182634/gemma-4-12B-it-Q4_K_M.gguf" } ] }, { "id": "Qwen/Qwen3-0.6B-Base", "name": "Qwen3-0.6B-Base BF16", "short_name": "Qwen3 0.6B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 950256, "tags": [ "transformers", "safetensors", "qwen3", "text-generation", "conversational", "arxiv:2505.09388", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 0.59604992, "active_params_b": 0.59604992, "layers": 28, "attention_heads": 16, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 1.19209984, "active_weight_gb": 1.19209984, "kv_alloc_gb_per_1k_tokens": 0.114688, "kv_read_gb_per_1k_tokens": 0.114688, "notes": "Manually audited on 2026-07-05 from HF API commit da87bfb608c14b7cf20ba1ce41287e8de496c0cd, served config, Qwen3 0.6B fine-tune comparison, and a direct safetensors header range read. The authoritative bounds profile uses exact stored BF16 tensor bytes: 1.19209984 GB resident and swept ordinary text traffic. The checkpoint stores no lm_head.weight, so model.embed_tokens.weight is charged as the tied output projection for ordinary decode." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-0.6B-Base" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-0.6B-Base/raw/da87bfb608c14b7cf20ba1ce41287e8de496c0cd/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-0.6B-Base" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/Qwen/Qwen3-0.6B-Base/resolve/da87bfb608c14b7cf20ba1ce41287e8de496c0cd/model.safetensors" } ] }, { "id": "DavidAU/Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking-NEO-CODE-Di-IMatrix-MAX-GGUF", "name": "Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking-NEO-CODE-Di-IMatrix-MAX-GGUF GGUF", "short_name": "Qwen3.6 40B Claude 4.6 Opus Deckard Heretic Unce", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 504420, "tags": [ "gguf", "unsloth", "fine tune", "heretic", "uncensored", "abliterated", "multi-stage tuned.", "all use cases", "coder", "creative", "creative writing", "fiction writing", "plot generation", "sub-plot generation", "story generation", "scene continue", "storytelling", "fiction story", "science fiction", "romance", "all genres", "story", "writing", "vivid prosing", "vivid writing", "fiction", "roleplaying", "bfloat16", "imatrix", "neo imatrix", "di-matrix", "image-text-to-text", "en", "zh", "dataset:TeichAI/claude-4.5-opus-high-reasoning-250x", "dataset:DavidAU/PkDick-Deckard-5-Datasets", "base_model:DavidAU/Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking", "base_model:quantized:DavidAU/Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen35 GGUF hybrid text", "total_params_b": 39.072596736, "active_params_b": 37.801198336, "layers": 96, "attention_heads": 24, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q8_0 HIGH", "weight_bytes_per_param": 1.0966049332606098, "resident_weight_gb": 42.847202336, "active_weight_gb": 41.485323264, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "kv_alloc_fixed_gb": 0.232390656, "kv_read_fixed_gb": 0.232390656, "notes": "Manually audited on 2026-07-05 from HF API commit 8f6654b98bf23b4d18374b7f72312cfba61d66db, model card, pinned full-precision base config, linked GGUF file sizes, and a GGUF tensor-index range read. The authoritative bounds profile uses the API-selected HIGH-Q8_0 artifact: 42.847202336 GB resident, 41.485323264 GB swept ordinary text traffic, 1.361879072 GB resident-only token embedding plus GGUF metadata/header/file overhead, and Qwen3.6 hybrid KV plus fixed DeltaNet state. Separate mmproj sidecars and Q4_K_M alias behavior are not included unless explicitly loaded." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/DavidAU/Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking-NEO-CODE-Di-IMatrix-MAX-GGUF" }, { "label": "Hugging Face 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" }, { "label": "Full-precision base config", "url": "https://huggingface.co/DavidAU/Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking/raw/bc23dff65597927a7f43e74b7a0deb6e49d773da/config.json" }, { "label": "Selected GGUF artifact", "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" } ] }, { "id": "Doradus-AI/RnJ-1-Instruct-FP8", "name": "RnJ-1-Instruct-FP8 FP8", "short_name": "RnJ 1 Instruct FP8", "license": "gemma", "hf_pipeline_tag": "text-generation", "hf_downloads": 527182, "tags": [ "transformers", "safetensors", "gemma3_text", "text-generation", "gemma3", "rnj", "doradus", "instruction-following", "fp8", "quantized", "vllm", "sglang", "conversational", "en", "license:gemma", "text-generation-inference", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma3_text (Gemma3ForCausalLM)", "total_params_b": 8.83734528, "active_params_b": 8.312008704, "layers": 32, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "compressed-tensors FP8 dynamic W8A8", "weight_bytes_per_param": 1.3573642386868472, "resident_weight_gb": 11.995496448, "active_weight_gb": 9.894150144, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "notes": "Manually audited on 2026-07-05 from HF API commit 3705c7b6c3d6b1591bd3d14a7ffa26b985eabe4d, served compressed-tensors config, model card, quantization recipe, safetensors index, linked-object HEAD checks, and direct safetensors header range reads. The authoritative bounds profile uses exact tensor payload bytes: 11.995496448 GB resident, 9.894150144 GB swept ordinary text traffic, and 2.101346304 GB resident-only input embedding. lm_head.weight is stored separately as F32 and remains swept." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Doradus-AI/RnJ-1-Instruct-FP8" }, { "label": "Model config", "url": "https://huggingface.co/Doradus-AI/RnJ-1-Instruct-FP8/raw/3705c7b6c3d6b1591bd3d14a7ffa26b985eabe4d/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Doradus-AI/RnJ-1-Instruct-FP8" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Doradus-AI/RnJ-1-Instruct-FP8/resolve/3705c7b6c3d6b1591bd3d14a7ffa26b985eabe4d/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen2.5-Coder-7B-Instruct-GPTQ-Int4", "name": "Qwen2.5-Coder-7B-Instruct-GPTQ-Int4 4-bit", "short_name": "Qwen2.5 Coder 7B Instruct GPTQ Int4", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 527068, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "code", "codeqwen", "chat", "qwen", "qwen-coder", "conversational", "en", "arxiv:2409.12186", "arxiv:2309.00071", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-Coder-7B-Instruct", "base_model:quantized:Qwen/Qwen2.5-Coder-7B-Instruct", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "4-bit", "gptq", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 7.6156, "active_params_b": 7.070619136, "layers": 28, "attention_heads": 28, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GPTQ Int4", "weight_bytes_per_param": 0.732084862626024, "resident_weight_gb": 5.575277568, "active_weight_gb": 4.485282816, "kv_alloc_gb_per_1k_tokens": 0.057344, "kv_read_gb_per_1k_tokens": 0.057344, "notes": "Manually audited on 2026-07-05 from HF API commit 3d0e91af138113d0b17d8011adf140d22423d13e, served GPTQ config, model card, audited BF16 base profile comparison, safetensors index, linked-object HEAD checks, and direct safetensors header range reads. The authoritative bounds profile uses exact tensor payload bytes: 5.575277568 GB resident, 4.485282816 GB swept ordinary text traffic, and 1.089994752 GB resident-only input embedding." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct-GPTQ-Int4" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct-GPTQ-Int4/raw/3d0e91af138113d0b17d8011adf140d22423d13e/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Coder-7B-Instruct-GPTQ-Int4" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct-GPTQ-Int4/raw/3d0e91af138113d0b17d8011adf140d22423d13e/model.safetensors.index.json" } ] }, { "id": "lukealonso/MiniMax-M2.7-NVFP4", "name": "MiniMax-M2.7-NVFP4 NVFP4", "short_name": "MiniMax M2.7 NVFP4", "license": "mit", "hf_pipeline_tag": "unknown", "hf_downloads": 518183, "tags": [ "safetensors", "minimax_m2", "custom_code", "base_model:MiniMaxAI/MiniMax-M2.7", "base_model:quantized:MiniMaxAI/MiniMax-M2.7", "license:mit", "8-bit", "modelopt", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "minimax_m2 (MiniMaxM2ForCausalLM)", "total_params_b": 228.703644928, "active_params_b": 10.41655168, "layers": 62, "attention_heads": 48, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 196608, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "NVFP4", "weight_bytes_per_param": 0.587667614175157, "resident_weight_gb": 134.401725368, "active_weight_gb": 10.73885657375, "kv_alloc_gb_per_1k_tokens": 0.253952, "kv_read_gb_per_1k_tokens": 0.253952, "notes": "Manually audited on 2026-07-05 from HF API commit db821d7a3ce29ee96d80a1cae88d878d8586b54e, model card, served ModelOpt config, hf_quant_config, audited FP8 base profile comparison, safetensors index, linked-object HEAD checks, and direct safetensors header range reads. Active traffic is fixed decode traffic plus 8 routed expert groups; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lukealonso/MiniMax-M2.7-NVFP4" }, { "label": "Model config", "url": "https://huggingface.co/lukealonso/MiniMax-M2.7-NVFP4/raw/db821d7a3ce29ee96d80a1cae88d878d8586b54e/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M2.7" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lukealonso/MiniMax-M2.7-NVFP4" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/lukealonso/MiniMax-M2.7-NVFP4/resolve/db821d7a3ce29ee96d80a1cae88d878d8586b54e/model.safetensors.index.json" } ] }, { "id": "google/gemma-4-E4B", "name": "gemma-4-E4B BF16", "short_name": "gemma 4 E4B", "license": "apache-2.0", "hf_pipeline_tag": "any-to-any", "hf_downloads": 532775, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "any-to-any", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 7.996157418, "active_params_b": 4.699496746, "layers": 42, "attention_heads": 8, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 15.992314836, "active_weight_gb": 9.398993492, "kv_alloc_gb_per_1k_tokens": 0.086016, "kv_read_gb_per_1k_tokens": 0.086016, "notes": "Manually audited on 2026-07-05 from current HF API metadata, model card, pinned served config, instruction-tuned profile/config comparison, linked-object HEAD check, and direct safetensors header range read. Bound weights use exact BF16 header bytes: 15.992314836 GB resident, 9.398993492 GB swept ordinary text traffic, and 6.593321344 GB resident-only audio/vision/per-layer-embedding package." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-4-E4B" }, { "label": "Model config", "url": "https://huggingface.co/google/gemma-4-E4B/raw/a24c9379fd3839ae84e97f0b6aa3152fce9bd033/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-E4B" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/google/gemma-4-E4B/resolve/a24c9379fd3839ae84e97f0b6aa3152fce9bd033/model.safetensors" } ] }, { "id": "QuantTrio/Qwen3.5-9B-AWQ", "name": "Qwen3.5-9B-AWQ 4-bit", "short_name": "Qwen3.5 9B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 453471, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "vLLM", "AWQ", "conversational", "base_model:Qwen/Qwen3.5-9B", "base_model:quantized:Qwen/Qwen3.5-9B", "license:apache-2.0", "endpoints_compatible", "4-bit", "awq", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 9.653104368, "active_params_b": 7.936684544, "layers": 32, "attention_heads": 16, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "AWQ INT4", "weight_bytes_per_param": 1.282113377021762, "resident_weight_gb": 12.37637424, "active_weight_gb": 8.943534592, "kv_alloc_gb_per_1k_tokens": 0.032768, "kv_read_gb_per_1k_tokens": 0.032768, "kv_alloc_fixed_gb": 0.051904512, "kv_read_fixed_gb": 0.051904512, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, pinned served AWQ config, Qwen3.5 9B base config comparison, model card, safetensors index, linked-object HEAD checks, direct shard header range reads, and the Transformers qwen3_5 implementation. Bound weights separate the exact 12.376374240 GB resident package from 8.943534592 GB ordinary text-decode swept traffic; the audited profile supersedes the generated ideal 4-bit estimate and adds the fixed DeltaNet runtime state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/QuantTrio/Qwen3.5-9B-AWQ" }, { "label": "Model config", "url": "https://huggingface.co/QuantTrio/Qwen3.5-9B-AWQ/raw/938f8e3ef86c9d1e9bec3705e149694c172592f1/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-9B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/QuantTrio/Qwen3.5-9B-AWQ" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/QuantTrio/Qwen3.5-9B-AWQ/raw/938f8e3ef86c9d1e9bec3705e149694c172592f1/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen2.5-7B", "name": "Qwen2.5-7B BF16", "short_name": "Qwen2.5 7B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 544225, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "conversational", "en", "arxiv:2407.10671", "license:apache-2.0", "eval-results", "text-generation-inference", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 7.615616512, "active_params_b": 7.070619136, "layers": 28, "attention_heads": 28, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 15.231233024, "active_weight_gb": 14.141238272, "kv_alloc_gb_per_1k_tokens": 0.057344, "kv_read_gb_per_1k_tokens": 0.057344, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, model card, pinned served config, safetensors index, linked-object HEAD checks, direct shard header range reads, and comparison against the existing Qwen2.5 7B Instruct profile. Bound weights separate the exact 15.231233024 GB resident BF16 package from 14.141238272 GB ordinary text-decode swept traffic." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-7B" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-7B/raw/d149729398750b98c0af14eb82c78cfe92750796/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-7B" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Qwen/Qwen2.5-7B/raw/d149729398750b98c0af14eb82c78cfe92750796/model.safetensors.index.json" } ] }, { "id": "nvidia/KVzap-mlp-Qwen3-8B", "name": "KVzap-mlp-Qwen3-8B F32", "short_name": "KVzap mlp Qwen3 8B", "license": "apache-2.0", "hf_pipeline_tag": "other", "hf_downloads": 693791, "tags": [ "transformers", "safetensors", "kvzap", "nvidia", "pytorch", "other", "dataset:nvidia/Nemotron-Pretraining-Dataset-sample", "arxiv:2506.05345", "arxiv:2601.07891", "arxiv:2505.23416", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "Auxiliary KV pruning model", "detail": "kvzap (KVzapModel)", "total_params_b": 0.075663648, "active_params_b": 0.075663648, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": 2048, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "F32", "weight_bytes_per_param": 4, "resident_weight_gb": 0.302654592, "active_weight_gb": 0.302654592, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manually reviewed on 2026-07-05 from current Hugging Face API metadata, model card, pinned KVzap config, linked-object HEAD check, and direct safetensors header range read. This is an auxiliary F32 KV-cache pruning scorer used with a separate LM through kvpress, not a standalone decoder-only language model; the profile intentionally fails closed for production throughput bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/KVzap-mlp-Qwen3-8B" }, { "label": "Model config", "url": "https://huggingface.co/nvidia/KVzap-mlp-Qwen3-8B/raw/bd5c5917846617da4311539859c137a262a6348b/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/KVzap-mlp-Qwen3-8B" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/nvidia/KVzap-mlp-Qwen3-8B/resolve/bd5c5917846617da4311539859c137a262a6348b/model.safetensors" } ] }, { "id": "Qwen/Qwen3.5-397B-A17B", "name": "Qwen3.5-397B-A17B BF16/F32", "short_name": "Qwen3.5 397B A17B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 420092, "tags": [ "transformers", "safetensors", "qwen3_5_moe", "image-text-to-text", "conversational", "license:apache-2.0", "eval-results", "endpoints_compatible", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_5_moe_text", "total_params_b": 403.397928944, "active_params_b": 16.331930816, "layers": 60, "attention_heads": 32, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "BF16/F32", "weight_bytes_per_param": 2, "resident_weight_gb": 806.795875168, "active_weight_gb": 32.663861632, "kv_alloc_gb_per_1k_tokens": 0.03072, "kv_read_gb_per_1k_tokens": 0.03072, "kv_alloc_fixed_gb": 0.19316736, "kv_read_fixed_gb": 0.19316736, "notes": "Manually audited on 2026-07-05 from current HF API metadata, model card, pinned served config, range-read safetensors shard headers, existing FP8/NVFP4 derivative profile comparisons, and the Transformers qwen3_5 implementation. Bound weights use exact stored bytes: 806.795875168 GB resident, 17.564367232 GB fixed text traffic, 1.509949440 GB per routed expert, 10 routed experts plus 1 shared expert, 15 full-attention layers, and a fixed DeltaNet recurrent-state charge of 0.19316736 GB per session/token." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3.5-397B-A17B" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3.5-397B-A17B/raw/8472618112abcbd45acbcdc58436aff4233c23f7/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-397B-A17B" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Qwen/Qwen3.5-397B-A17B/resolve/8472618112abcbd45acbcdc58436aff4233c23f7/model.safetensors.index.json" } ] }, { "id": "nvidia/Nemotron-Labs-Diffusion-8B-Base", "name": "Nemotron-Labs-Diffusion-8B-Base BF16", "short_name": "Nemotron Labs Diffusion 8B", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 479470, "tags": [ "transformers", "safetensors", "nemotron_labs_diffusion", "feature-extraction", "nvidia", "pytorch", "text-generation", "conversational", "custom_code", "license:other", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "nemotron_labs_diffusion tri-mode decoder", "total_params_b": 8.48955392, "active_params_b": 7.952683008, "layers": 34, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 4096, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 16.97910784, "active_weight_gb": 15.905366016, "kv_alloc_gb_per_1k_tokens": 0.139264, "kv_read_gb_per_1k_tokens": 0.139264, "notes": "Manual correction on 2026-07-05 from live Hugging Face API metadata, pinned config/generation config, model card, custom modeling code, and direct safetensors header range read. The package has 16.979107840 GB BF16 resident tensors; AR-mode non-embedding swept bytes are 15.905366016 GB. Production bounds are disabled in the model profile because the repo exposes AR, diffusion, and self-speculation modes without a comparison-time generation-mode selector." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-8B-Base" }, { "label": "Model config", "url": "https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-8B-Base/raw/59ff0ffee284112fc6ccf37493ced41a11031434/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/Nemotron-Labs-Diffusion-8B-Base" }, { "label": "Pinned safetensors header audit", "url": "https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-8B-Base/resolve/59ff0ffee284112fc6ccf37493ced41a11031434/model.safetensors" } ] }, { "id": "Qwen/Qwen2.5-3B", "name": "Qwen2.5-3B BF16", "short_name": "Qwen2.5 3B", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 493002, "tags": [ "safetensors", "qwen2", "text-generation", "conversational", "en", "arxiv:2407.10671", "license:other", "eval-results", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 3.085938688, "active_params_b": 3.085938688, "layers": 36, "attention_heads": 16, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 6.171877376, "active_weight_gb": 6.171877376, "kv_alloc_gb_per_1k_tokens": 0.036864, "kv_read_gb_per_1k_tokens": 0.036864, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, model card, pinned served config, safetensors index, linked-object HEAD checks, direct shard header range reads, and comparison against the existing Qwen2.5 3B Instruct profile. Bound weights use exact BF16 tensor payload." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-3B" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-3B/raw/3aab1f1954e9cc14eb9509a215f9e5ca08227a9b/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-3B" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Qwen/Qwen2.5-3B/raw/3aab1f1954e9cc14eb9509a215f9e5ca08227a9b/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen3-30B-A3B-FP8", "name": "Qwen3-30B-A3B-FP8 FP8", "short_name": "Qwen3 30B A3B FP8", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 508657, "tags": [ "transformers", "safetensors", "text-generation", "arxiv:2309.00071", "arxiv:2505.09388", "base_model:Qwen/Qwen3-30B-A3B", "base_model:finetune:Qwen/Qwen3-30B-A3B", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_moe (Qwen3MoeForCausalLM)", "total_params_b": 30.533947392, "active_params_b": 3, "layers": 48, "attention_heads": 32, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "FP8", "weight_bytes_per_param": 1, "resident_weight_gb": 32.444792832, "active_weight_gb": 4.01440768, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, model card, pinned served FP8 config, audited BF16 profile comparison, safetensors index, linked-object HEAD checks, and direct shard header range reads. Resident tensor payload is 32.444792832 GB across F8_E4M3 and F32 tensors. Ordinary single-session MoE text-decode traffic is 4.014407680 GB: 2.202025984 GB fixed traffic plus 8 routed expert groups at 0.226547712 GB each." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-FP8" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-FP8/raw/d206ba732169f29bb77fbf80fc2c4b81d4d30782/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-30B-A3B-FP8" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-FP8/raw/d206ba732169f29bb77fbf80fc2c4b81d4d30782/model.safetensors.index.json" } ] }, { "id": "RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8", "name": "Meta-Llama-3.1-8B-Instruct-FP8 FP8", "short_name": "Meta Llama 3.1 8B Instruct FP8", "license": "llama3.1", "hf_pipeline_tag": "text-generation", "hf_downloads": 534098, "tags": [ "transformers", "safetensors", "llama", "text-generation", "fp8", "vllm", "conversational", "en", "de", "fr", "it", "pt", "hi", "es", "th", "base_model:meta-llama/Llama-3.1-8B-Instruct", "base_model:quantized:meta-llama/Llama-3.1-8B-Instruct", "license:llama3.1", "text-generation-inference", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "llama (LlamaForCausalLM)", "total_params_b": 8.030261696, "active_params_b": 7.50492512, "layers": 32, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "FP8", "weight_bytes_per_param": 1, "resident_weight_gb": 9.081201536, "active_weight_gb": 8.030528384, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, model card, served static FP8 config, recipe.yaml, safetensors index, linked-object HEAD checks, and direct shard header range reads. Bound weights separate the 9.081201536 GB resident compressed-tensors package from the 8.030528384 GB ordinary text-decode swept subset; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8" }, { "label": "Model config", "url": "https://huggingface.co/RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8/raw/12fd6884d2585dd4d020373e7f39f74507b31866/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8/raw/12fd6884d2585dd4d020373e7f39f74507b31866/model.safetensors.index.json" } ] }, { "id": "deepseek-ai/DeepSeek-R1-Distill-Qwen-14B", "name": "DeepSeek-R1-Distill-Qwen-14B BF16", "short_name": "DeepSeek R1 Distill Qwen 14B", "license": "mit", "hf_pipeline_tag": "unknown", "hf_downloads": 452833, "tags": [ "transformers", "safetensors", "arxiv:2501.12948", "license:mit", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 14.770033664, "active_params_b": 13.991465984, "layers": 48, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 29.540067328, "active_weight_gb": 27.982931968, "kv_alloc_gb_per_1k_tokens": 0.196608, "kv_read_gb_per_1k_tokens": 0.196608, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, pinned model card, pinned served config, Qwen2.5 14B base config comparison, safetensors index, and direct shard header range reads. Bound weights separate the exact 29.540067328 GB resident package from 27.982931968 GB ordinary text-decode swept traffic; the audited profile supersedes the generated flat dense estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B" }, { "label": "Model config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B/raw/1df8507178afcc1bef68cd8c393f61a886323761/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B" }, { "label": "Pinned base config", "url": "https://huggingface.co/Qwen/Qwen2.5-14B/raw/97e1e76335b7017d8f67c08a19d103c0504298c9/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B/raw/1df8507178afcc1bef68cd8c393f61a886323761/model.safetensors.index.json" } ] }, { "id": "cyankiwi/Qwen3.6-27B-AWQ-BF16-INT4", "name": "Qwen3.6-27B-AWQ-BF16-INT4 4-bit", "short_name": "Qwen3.6 27B AWQ BF16 INT4", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 570671, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "conversational", "base_model:Qwen/Qwen3.6-27B", "base_model:quantized:Qwen/Qwen3.6-27B", "license:apache-2.0", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 27.781427952, "active_params_b": 25.624600064, "layers": 64, "attention_heads": 24, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "compressed-tensors BF16/INT4", "weight_bytes_per_param": 1.0192369723011852, "resident_weight_gb": 28.315858512, "active_weight_gb": 24.531487744, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, pinned served compressed-tensors config, Qwen3.6 27B base config comparison, model card, safetensors index, direct shard header range reads, comparison against the existing AWQ-INT4 profile, and the Transformers qwen3_5 implementation. Bound weights separate the exact 28.315858512 GB resident package from 24.531487744 GB ordinary text-decode swept traffic; the audited profile supersedes the generated ideal 4-bit estimate and adds the fixed DeltaNet runtime state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/cyankiwi/Qwen3.6-27B-AWQ-BF16-INT4" }, { "label": "Model config", "url": "https://huggingface.co/cyankiwi/Qwen3.6-27B-AWQ-BF16-INT4/raw/a3d9607a84817d20ed404d0d12765e70c8a2026f/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.6-27B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3.6-27B-AWQ-BF16-INT4" }, { "label": "Pinned base config", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/cyankiwi/Qwen3.6-27B-AWQ-BF16-INT4/resolve/a3d9607a84817d20ed404d0d12765e70c8a2026f/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen2.5-Math-1.5B-Instruct", "name": "Qwen2.5-Math-1.5B-Instruct BF16", "short_name": "Qwen2.5 Math 1.5B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 484721, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "chat", "conversational", "en", "arxiv:2409.12122", "base_model:Qwen/Qwen2.5-Math-1.5B", "base_model:finetune:Qwen/Qwen2.5-Math-1.5B", "license:apache-2.0", "eval-results", "text-generation-inference", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 1.543714304, "active_params_b": 1.543714304, "layers": 28, "attention_heads": 12, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 4096, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 3.087428608, "active_weight_gb": 3.087428608, "kv_alloc_gb_per_1k_tokens": 0.028672, "kv_read_gb_per_1k_tokens": 0.028672, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, pinned served config, base config comparison, and a direct safetensors header range read. downloads=484721; data quality: config, API params, safetensors header, base:Qwen/Qwen2.5-Math-1.5B. Bound weights use exact BF16 tensor payload." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Math-1.5B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-Math-1.5B-Instruct/raw/aafeb0fc6f22cbf0eaeed126eff8be45b0360a35/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Math-1.5B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Math-1.5B-Instruct" }, { "label": "Safetensors header", "url": "https://huggingface.co/Qwen/Qwen2.5-Math-1.5B-Instruct/resolve/aafeb0fc6f22cbf0eaeed126eff8be45b0360a35/model.safetensors" } ] }, { "id": "RedHatAI/gemma-3-27b-it-FP8-dynamic", "name": "gemma-3-27b-it-FP8-dynamic FP8", "short_name": "gemma 3 27b it FP8 dynamic", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 520563, "tags": [ "transformers", "safetensors", "gemma3", "image-text-to-text", "vllm", "vision", "fp8", "conversational", "en", "base_model:google/gemma-3-27b-it", "base_model:quantized:google/gemma-3-27b-it", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense multimodal transformer", "detail": "gemma3_text (Gemma3ForConditionalGeneration)", "total_params_b": 27.436247664, "active_params_b": 27.013187328, "layers": 62, "attention_heads": 32, "kv_heads": 16, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "compressed-tensors FP8 dynamic W8A8", "weight_bytes_per_param": 1.0669875154397133, "resident_weight_gb": 29.274133728, "active_weight_gb": 28.428013056, "kv_alloc_gb_per_1k_tokens": 0.08192, "kv_read_gb_per_1k_tokens": 0.08192, "notes": "Manually audited on 2026-07-05 from current HF API metadata, served compressed-tensors config, model card metadata, safetensors index, direct safetensors header range reads, and Gemma 3 local/global attention documentation. downloads=520563. Exact tensor payload bytes: 29.274133728 GB resident, 28.428013056 GB swept ordinary text traffic, and 0.846120672 GB resident-only vision/projector bytes. KV fields show only the full-context component per 1K tokens; the audited profile also charges 52 sliding-window layers capped at 1024 tokens." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/RedHatAI/gemma-3-27b-it-FP8-dynamic" }, { "label": "Model config", "url": "https://huggingface.co/RedHatAI/gemma-3-27b-it-FP8-dynamic/raw/306078afe860ef87821018f28078ecf762f31455/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-3-27b-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/RedHatAI/gemma-3-27b-it-FP8-dynamic" }, { "label": "Safetensors index", "url": "https://huggingface.co/RedHatAI/gemma-3-27b-it-FP8-dynamic/raw/306078afe860ef87821018f28078ecf762f31455/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen3-8B-AWQ", "name": "Qwen3-8B-AWQ 4-bit", "short_name": "Qwen3 8B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 378299, "tags": [ "transformers", "safetensors", "qwen3", "text-generation", "conversational", "arxiv:2309.00071", "arxiv:2505.09388", "base_model:Qwen/Qwen3-8B", "base_model:quantized:Qwen/Qwen3-8B", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "4-bit", "awq", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 8.19073536, "active_params_b": 7.568405504, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "AWQ 4-bit GEMM", "weight_bytes_per_param": 0.744558191170762, "resident_weight_gb": 6.098479104, "active_weight_gb": 4.853819392, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, served AWQ config, audited BF16 base config/profile comparison, safetensors index, and direct safetensors header range reads. downloads=378299. Exact direct tensor payload bytes: 6.098479104 GB resident, 4.853819392 GB swept ordinary text traffic, and 1.244659712 GB resident-only input embedding. The audited profile supersedes the generated ideal 4-bit estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-8B-AWQ" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-8B-AWQ/raw/4da05a8edb55c6046cce958586c33b61da07bb79/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-8B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-8B-AWQ" }, { "label": "Safetensors index", "url": "https://huggingface.co/Qwen/Qwen3-8B-AWQ/raw/4da05a8edb55c6046cce958586c33b61da07bb79/model.safetensors.index.json" } ] }, { "id": "nvidia/Llama-3_3-Nemotron-Super-49B-v1_5", "name": "Llama-3_3-Nemotron-Super-49B-v1_5 BF16", "short_name": "Llama 3 3 Nemotron Super 49B v1 5", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 259026, "tags": [ "transformers", "safetensors", "nemotron-nas", "text-generation", "nvidia", "llama-3", "pytorch", "conversational", "custom_code", "en", "arxiv:2411.19146", "arxiv:2505.00949", "arxiv:2502.00203", "license:other", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "nemotron-nas (DeciLMForCausalLM)", "total_params_b": 49.867145216, "active_params_b": 49.867145216, "layers": 80, "attention_heads": 64, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 99.734290432, "active_weight_gb": 97.632944128, "kv_alloc_gb_per_1k_tokens": 0.200704, "kv_read_gb_per_1k_tokens": 0.200704, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, pinned served DeciLM config, custom runtime code, block config dataclasses, safetensors index, and direct safetensors shard header range reads. downloads=259026. Exact BF16 tensor payload bytes: 99.734290432 GB resident, 97.632944128 GB swept ordinary text traffic, and 2.101346304 GB resident-only input embedding. The audited DeciLM profile supersedes the generated estimate, which incorrectly charged 64 KV heads across all 80 layers." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1_5" }, { "label": "Model config", "url": "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1_5/raw/420ba7d28211abf116b8b103ab700d92619daf98/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/Llama-3_3-Nemotron-Super-49B-v1_5" }, { "label": "Custom DeciLM runtime", "url": "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1_5/raw/420ba7d28211abf116b8b103ab700d92619daf98/modeling_decilm.py" }, { "label": "Safetensors index", "url": "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1_5/raw/420ba7d28211abf116b8b103ab700d92619daf98/model.safetensors.index.json" } ] }, { "id": "unsloth/gemma-4-12B-it-qat-GGUF", "name": "gemma-4-12B-it-qat-GGUF GGUF 4-bit", "short_name": "gemma 4 12B it qat", "license": "apache-2.0", "hf_pipeline_tag": "any-to-any", "hf_downloads": 509756, "tags": [ "transformers", "gguf", "gemma4", "unsloth", "gemma", "google", "any-to-any", "base_model:google/gemma-4-12B-it-qat-q4_0-unquantized", "base_model:quantized:google/gemma-4-12B-it-qat-q4_0-unquantized", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_unified_text", "total_params_b": 11.907350576, "active_params_b": 11.907350576, "layers": 48, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF UD-Q4_K_XL QAT", "weight_bytes_per_param": 0.5640511955310384, "resident_weight_gb": 6.716355328, "active_weight_gb": 6.700533248, "kv_alloc_gb_per_1k_tokens": 0.335872, "kv_read_gb_per_1k_tokens": 0.335872, "notes": "Manually audited on 2026-07-05 from HF API commit 7102bdea62863acff919c945405ef29973113d66, Google QAT unquantized base config, model card, linked GGUF file sizes, and a GGUF tensor-index range read. The authoritative bounds profile uses the selected gemma-4-12B-it-qat-UD-Q4_K_XL.gguf artifact: 6.716355328 GB resident, 6.700533248 GB swept text traffic, 0.01582208 GB resident-only GGUF metadata/header bytes, and layered Gemma 4 hybrid KV. The separate mmproj and MTP GGUF files are not included unless explicitly loaded. Legacy per-1k KV fields show the short-context unsaturated coefficient; long-context calculation must use the profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/gemma-4-12B-it-qat-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/google/gemma-4-12B-it-qat-q4_0-unquantized/raw/58540658b6c08edab2ddc1fbde7f28cc9987ced3/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-12B-it-qat-q4_0-unquantized" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/gemma-4-12B-it-qat-GGUF" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/unsloth/gemma-4-12B-it-qat-GGUF/resolve/7102bdea62863acff919c945405ef29973113d66/gemma-4-12B-it-qat-UD-Q4_K_XL.gguf" } ] }, { "id": "cyankiwi/Qwen3.5-35B-A3B-AWQ-4bit", "name": "Qwen3.5-35B-A3B-AWQ-4bit 4-bit", "short_name": "Qwen3.5 35B A3B AWQ 4bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 441851, "tags": [ "transformers", "safetensors", "qwen3_5_moe", "image-text-to-text", "conversational", "base_model:Qwen/Qwen3.5-35B-A3B", "base_model:quantized:Qwen/Qwen3.5-35B-A3B", "license:apache-2.0", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_5_moe_text", "total_params_b": 36.983684464, "active_params_b": 3, "layers": 40, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "compressed-tensors AWQ Int4 packed I32 + BF16/I64 side tensors", "weight_bytes_per_param": 0.6608699351139912, "resident_weight_gb": 24.441405152, "active_weight_gb": 4.445839616, "kv_alloc_gb_per_1k_tokens": 0.02048, "kv_read_gb_per_1k_tokens": 0.02048, "kv_alloc_fixed_gb": 0.06488064, "kv_read_fixed_gb": 0.06488064, "expert_param_b": 0.10710246636925808, "fixed_param_b": 5.870433817405875, "shared_expert_param_b": 0, "notes": "Manually corrected on 2026-07-05 from current Hugging Face API metadata, pinned compressed-tensors config, BF16 base config comparison, recipe.yaml, safetensors index, and direct range-read safetensors shard headers. Legacy fallback values mirror the audited profile: exact stored resident bytes, fixed always-on hybrid/shared-expert traffic, expected distinct routed experts, BF16 attention KV, and fixed BF16-convolution/F32-SSM DeltaNet recurrent state. The authoritative profile supersedes the generated 18.4918 GB resident estimate and the missing shared-expert count." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/cyankiwi/Qwen3.5-35B-A3B-AWQ-4bit" }, { "label": "Model config", "url": "https://huggingface.co/cyankiwi/Qwen3.5-35B-A3B-AWQ-4bit/raw/affb52d39eeb1c4d702372be794c7c2b36c32d47/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3.5-35B-A3B-AWQ-4bit" }, { "label": "Safetensors index", "url": "https://huggingface.co/cyankiwi/Qwen3.5-35B-A3B-AWQ-4bit/resolve/affb52d39eeb1c4d702372be794c7c2b36c32d47/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen3-8B-Base", "name": "Qwen3-8B-Base BF16", "short_name": "Qwen3 8B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 428867, "tags": [ "transformers", "safetensors", "qwen3", "text-generation", "conversational", "arxiv:2505.09388", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 8.19073536, "active_params_b": 7.568405504, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 16.38147072, "active_weight_gb": 15.136811008, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Manually corrected on 2026-07-05 from current Hugging Face API metadata, pinned served config, model card, safetensors index, direct range-read safetensors shard headers, and Qwen3 8B fine-tune comparison. Legacy fallback values mirror the audited profile: 16.381470720 GB resident BF16 tensor payload, 15.136811008 GB ordinary text swept traffic excluding the input embedding lookup, and full-context BF16 KV. The authoritative profile supersedes the generated rounded dense estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-8B-Base" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-8B-Base/raw/49e3418fbbbca6ecbdf9608b4d22e5a407081db4/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-8B-Base" }, { "label": "Safetensors index", "url": "https://huggingface.co/Qwen/Qwen3-8B-Base/resolve/49e3418fbbbca6ecbdf9608b4d22e5a407081db4/model.safetensors.index.json" } ] }, { "id": "moonshotai/Kimi-K2-Instruct", "name": "Kimi-K2-Instruct FP8", "short_name": "Kimi K2", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 379918, "tags": [ "transformers", "safetensors", "kimi_k2", "text-generation", "conversational", "custom_code", "doi:10.57967/hf/5976", "license:other", "eval-results", "endpoints_compatible", "fp8", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "kimi_k2 (DeepseekV3ForCausalLM)", "total_params_b": 1026.470731056, "active_params_b": 32, "layers": 61, "attention_heads": 64, "kv_heads": 64, "head_dim": 112, "max_context_tokens": 131072, "routed_experts": 384, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "FP8 F8_E4M3 + BF16/F32 side tensors", "weight_bytes_per_param": 1.0026328326587741, "resident_weight_gb": 1029.17325672, "active_weight_gb": 33.03514984, "kv_alloc_gb_per_1k_tokens": 2.49856, "kv_read_gb_per_1k_tokens": 2.49856, "expert_param_b": 2.6361161872099905, "shared_expert_param_b": 0, "fixed_param_b": 11.859472712925568, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, served FP8 config, model card, deployment guide, custom modeling code, safetensors index, and direct safetensors header range reads. Bound weights separate the 1029.173256720 GB resident block-FP8 package from the 33.035149840 GB ordinary text-decode active traffic; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/moonshotai/Kimi-K2-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/moonshotai/Kimi-K2-Instruct/raw/main/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/moonshotai/Kimi-K2-Instruct" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/moonshotai/Kimi-K2-Instruct/resolve/fd1984e2b7a3350dbf7305fe73a4ede25c14de50/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen2.5-32B-Instruct-GPTQ-Int4", "name": "Qwen2.5-32B-Instruct-GPTQ-Int4 4-bit", "short_name": "Qwen2.5 32B Instruct GPTQ Int4", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 463895, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "chat", "conversational", "en", "arxiv:2309.00071", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-32B-Instruct", "base_model:quantized:Qwen/Qwen2.5-32B-Instruct", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "4-bit", "gptq", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 32.763876352, "active_params_b": 31.985308672, "layers": 64, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GPTQ 4-bit + F16 side tensors", "weight_bytes_per_param": 0.590398610475137, "resident_weight_gb": 19.343747072, "active_weight_gb": 17.786611712, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, served GPTQ config, model card, current base config comparison, safetensors index, and direct safetensors header range reads. Bound weights separate the 19.343747072 GB resident GPTQ package from the 17.786611712 GB ordinary text-decode swept subset; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct-GPTQ-Int4" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct-GPTQ-Int4/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-32B-Instruct-GPTQ-Int4" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct-GPTQ-Int4/resolve/c83e67dfb2664f5039fd4cd99e206799e27dd800/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen3-VL-4B-Instruct-FP8", "name": "Qwen3-VL-4B-Instruct-FP8 FP8", "short_name": "Qwen3 VL 4B Instruct FP8", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 474505, "tags": [ "transformers", "safetensors", "qwen3_vl", "image-text-to-text", "conversational", "arxiv:2505.09388", "arxiv:2502.13923", "arxiv:2409.12191", "arxiv:2308.12966", "base_model:Qwen/Qwen3-VL-4B-Instruct", "base_model:quantized:Qwen/Qwen3-VL-4B-Instruct", "license:apache-2.0", "endpoints_compatible", "fp8", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_vl_text", "total_params_b": 4.826993728, "active_params_b": 4.022689856, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "FP8 F8_E4M3 + BF16/F32 side tensors", "weight_bytes_per_param": 1.2473840811255348, "resident_weight_gb": 6.021115136, "active_weight_gb": 4.412507392, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, served FP8 config, model card, BF16 base profile comparison, safetensors index, and direct safetensors header range reads. Bound weights separate the 6.021115136 GB resident FP8/BF16/F32 multimodal package from the 4.412507392 GB ordinary text-decode swept subset; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct-FP8" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct-FP8/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-VL-4B-Instruct-FP8" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct-FP8/resolve/fefbb44cbcce8d1bb7e20b920b94f77432b3446d/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen3-Next-80B-A3B-Instruct-FP8", "name": "Qwen3-Next-80B-A3B-Instruct-FP8 FP8", "short_name": "Qwen3 Next 80B A3B Instruct FP8", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 476230, "tags": [ "transformers", "safetensors", "qwen3_next", "text-generation", "conversational", "arxiv:2309.00071", "arxiv:2404.06654", "arxiv:2505.09388", "arxiv:2501.15383", "base_model:Qwen/Qwen3-Next-80B-A3B-Instruct", "base_model:quantized:Qwen/Qwen3-Next-80B-A3B-Instruct", "license:apache-2.0", "endpoints_compatible", "fp8", "deploy:azure", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_next (Qwen3NextForCausalLM)", "total_params_b": 81.329784384, "active_params_b": 3.9322496, "layers": 48, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "FP8 F8_E4M3 + BF16/F32 side tensors", "weight_bytes_per_param": 1.008644861674296, "resident_weight_gb": 82.03286912, "active_weight_gb": 3.9322496, "kv_alloc_gb_per_1k_tokens": 0.024576, "kv_read_gb_per_1k_tokens": 0.024576, "kv_alloc_fixed_gb": 0.077856768, "kv_read_fixed_gb": 0.077856768, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, served FP8 config, model card, BF16 base config comparison, safetensors index, and direct safetensors header range reads. Bound weights separate the 82.032869120 GB resident FP8/BF16/F32 package from the 3.932249600 GB ordinary text-decode swept subset; top-level MTP tensors are resident-only for non-speculative decode. The audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct-FP8" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct-FP8/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-Next-80B-A3B-Instruct-FP8" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct-FP8/resolve/c5f5f263bdd5cc134092897864e8905d8fe7b928/model.safetensors.index.json" } ] }, { "id": "palmfuture/Qwen3.6-35B-A3B-GPTQ-Int4", "name": "Qwen3.6-35B-A3B-GPTQ-Int4 4-bit", "short_name": "Qwen3.6 35B A3B GPTQ Int4", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 449506, "tags": [ "transformers", "safetensors", "qwen3_5_moe", "image-text-to-text", "quantized", "gptq", "int4", "moe", "qwen3.6", "mtp", "speculative-decoding", "conversational", "en", "th", "zh", "multilingual", "base_model:Qwen/Qwen3.6-35B-A3B", "base_model:quantized:Qwen/Qwen3.6-35B-A3B", "license:apache-2.0", "endpoints_compatible", "4-bit", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_5_moe_text", "total_params_b": 35.951822704, "active_params_b": 4.408468736, "layers": 40, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "GPTQ Int4 + BF16/F16 side tensors", "weight_bytes_per_param": 0.6787737247403844, "resident_weight_gb": 24.403152608, "active_weight_gb": 4.408468736, "kv_alloc_gb_per_1k_tokens": 0.01024, "kv_read_gb_per_1k_tokens": 0.01024, "kv_alloc_fixed_gb": 0.06488064, "kv_read_fixed_gb": 0.06488064, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, served GPTQ config, model card, BF16 base config comparison, safetensors index, and direct safetensors header range reads. Bound weights separate the 24.403152608 GB resident GPTQ/BF16/F16 multimodal package from the 4.408468736 GB ordinary text-decode swept subset; MTP, vision, and input embedding tensors are resident-only for non-speculative decode. The audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/palmfuture/Qwen3.6-35B-A3B-GPTQ-Int4" }, { "label": "Model config", "url": "https://huggingface.co/palmfuture/Qwen3.6-35B-A3B-GPTQ-Int4/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/palmfuture/Qwen3.6-35B-A3B-GPTQ-Int4" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/palmfuture/Qwen3.6-35B-A3B-GPTQ-Int4/resolve/d1fef185160f938fca00c3c664f21250dd544d63/model.safetensors.index.json" } ] }, { "id": "QuantTrio/DeepSeek-V3.2-AWQ", "name": "DeepSeek-V3.2-AWQ 4-bit", "short_name": "DeepSeek V3.2", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 431262, "tags": [ "transformers", "safetensors", "deepseek_v32", "text-generation", "vLLM", "AWQ", "conversational", "base_model:deepseek-ai/DeepSeek-V3.2", "base_model:quantized:deepseek-ai/DeepSeek-V3.2", "license:mit", "endpoints_compatible", "4-bit", "awq", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "deepseek_v32 (DeepseekV32ForCausalLM)", "total_params_b": 685.355329792, "active_params_b": 22.898430208, "layers": 61, "attention_heads": 128, "kv_heads": 128, "head_dim": 56, "max_context_tokens": 163840, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "AWQ Int4 + F16/F32 side tensors", "weight_bytes_per_param": 0.5280610534069046, "resident_weight_gb": 361.909457408, "active_weight_gb": 22.898430208, "kv_alloc_gb_per_1k_tokens": 0.048068, "kv_read_gb_per_1k_tokens": 0.048068, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, served AWQ config, model card, official base config and inference comparison, safetensors index, and direct safetensors header range reads. Bound weights separate the 361.909457408 GB resident AWQ/F16/F32 package from the 22.898430208 GB ordinary text-decode swept subset; the next-token-prediction layer is resident-only for ordinary decode. The audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/QuantTrio/DeepSeek-V3.2-AWQ" }, { "label": "Model config", "url": "https://huggingface.co/QuantTrio/DeepSeek-V3.2-AWQ/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.2" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/QuantTrio/DeepSeek-V3.2-AWQ" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/QuantTrio/DeepSeek-V3.2-AWQ/resolve/340023cb6036c97c5c664ac944300e9d2b1a3f2e/model.safetensors.index.json" } ] }, { "id": "cyankiwi/Qwen3.5-2B-AWQ-4bit", "name": "Qwen3.5-2B-AWQ-4bit 4-bit", "short_name": "Qwen3.5 2B AWQ 4bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 465061, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "conversational", "base_model:Qwen/Qwen3.5-2B", "base_model:quantized:Qwen/Qwen3.5-2B", "license:apache-2.0", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 2.274069824, "active_params_b": 1.881825088, "layers": 24, "attention_heads": 8, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "compressed-tensors AWQ INT4/BF16", "weight_bytes_per_param": 1.0998742033349278, "resident_weight_gb": 2.501190736, "active_weight_gb": 1.792067552, "kv_alloc_gb_per_1k_tokens": 0.012288, "kv_read_gb_per_1k_tokens": 0.012288, "kv_alloc_fixed_gb": 0.019759104, "kv_read_fixed_gb": 0.019759104, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, pinned compressed-tensors config, Qwen3.5 2B base config comparison, model card, safetensors index, direct shard header range read, and the Transformers qwen3_5 implementation. Bound weights separate the exact 2.501190736 GB resident package from 1.792067552 GB ordinary text-decode swept traffic; the audited profile supersedes the generated ideal 4-bit estimate and adds the fixed DeltaNet runtime state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/cyankiwi/Qwen3.5-2B-AWQ-4bit" }, { "label": "Model config", "url": "https://huggingface.co/cyankiwi/Qwen3.5-2B-AWQ-4bit/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-2B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3.5-2B-AWQ-4bit" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/cyankiwi/Qwen3.5-2B-AWQ-4bit/resolve/e27b458e4c295748030ed367efaee3b2f2ccf32f/model.safetensors.index.json" } ] }, { "id": "deepseek-ai/DeepSeek-V2-Lite", "name": "DeepSeek-V2-Lite BF16", "short_name": "DeepSeek V2 Lite", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 483565, "tags": [ "transformers", "safetensors", "deepseek_v2", "text-generation", "conversational", "custom_code", "arxiv:2405.04434", "license:other", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "deepseek_v2 (DeepseekV2ForCausalLM)", "total_params_b": 15.706484224, "active_params_b": 2.451435008, "layers": 27, "attention_heads": 16, "kv_heads": 16, "head_dim": 128, "max_context_tokens": 163840, "routed_experts": 64, "routed_experts_per_token": 6, "shared_experts_per_token": 2 }, "adapter": { "kind": "moe", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 31.412968448, "active_weight_gb": 4.902870016, "kv_alloc_gb_per_1k_tokens": 0.27648, "kv_read_gb_per_1k_tokens": 0.27648, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, model card, pinned config, pinned custom modeling code, safetensors index, direct range-read shard headers, and byte-level comparison with the audited DeepSeek-V2-Lite-Chat profile. Bound weights separate the 31.412968448 GB resident BF16 checkpoint from 4.902870016 GB single-token active ordinary decode traffic; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite" }, { "label": "Model config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite/raw/604d5664dddd88a0433dbae533b7fe9472482de0/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-V2-Lite" }, { "label": "Pinned custom modeling code", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite/raw/604d5664dddd88a0433dbae533b7fe9472482de0/modeling_deepseek.py" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite/resolve/604d5664dddd88a0433dbae533b7fe9472482de0/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen2.5-VL-32B-Instruct", "name": "Qwen2.5-VL-32B-Instruct BF16/F32", "short_name": "Qwen2.5 VL 32B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 446745, "tags": [ "transformers", "safetensors", "multimodal", "image-text-to-text", "en", "arxiv:2309.00071", "arxiv:2502.13923", "license:apache-2.0", "eval-results", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2_5_vl (Qwen2_5_VLForConditionalGeneration)", "total_params_b": 33.452718336, "active_params_b": 31.985308672, "layers": 64, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 128000, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16/F32", "weight_bytes_per_param": 2.0411830199914553, "resident_weight_gb": 68.28312064, "active_weight_gb": 63.970617344, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, pinned config, model card, safetensors index, direct range-read shard headers, and comparison with the audited Qwen2.5 32B text profile. Bound weights separate the exact 68.283120640 GB resident BF16/F32 multimodal package from 63.970617344 GB ordinary text-decode swept traffic; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-32B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-32B-Instruct/raw/7cfb30d71a1f4f49a57592323337a4a4727301da/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-VL-32B-Instruct" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-32B-Instruct/resolve/7cfb30d71a1f4f49a57592323337a4a4727301da/model.safetensors.index.json" } ] }, { "id": "nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4", "name": "NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4 NVFP4", "short_name": "NVIDIA Nemotron 3 Ultra 550B A55B NVFP4", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 489280, "tags": [ "transformers", "safetensors", "nemotron_h", "text-generation", "nvidia", "pytorch", "nemotron-3", "latent-moe", "mtp", "conversational", "en", "fr", "es", "it", "de", "pt", "ja", "ko", "hi", "ar", "zh", "he", "dataset:nvidia/nemotron-post-training-v3", "dataset:nvidia/nemotron-pre-training-datasets", "license:other", "eval-results", "endpoints_compatible", "8-bit", "modelopt", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "nemotron_h (NemotronHForCausalLM)", "total_params_b": 550, "active_params_b": 55, "layers": 108, "attention_heads": 64, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 512, "routed_experts_per_token": 22, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "NVFP4/FP8/BF16", "weight_bytes_per_param": 0.5, "resident_weight_gb": 352.28406128, "active_weight_gb": 50.251769952, "kv_alloc_gb_per_1k_tokens": 0.006144, "kv_read_gb_per_1k_tokens": 0.006144, "kv_alloc_fixed_gb": 0.20840448, "kv_read_fixed_gb": 0.20840448, "expert_param_b": 1.132463616, "fixed_param_b": 75.589340352, "shared_expert_param_b": 0, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, pinned served config, BF16 config comparison, ModelOpt quantization config, model card serving notes, Nemotron-H runtime cache review, safetensors index, and direct range-read safetensors shard headers. Legacy fallback values mirror the audited profile: exact stored resident bytes, fixed always-on hybrid/shared-expert traffic, expected distinct routed experts, FP8 attention KV, and fixed float16-sized Mamba recurrent state. The authoritative profile supersedes the generated 167.5194 GB resident estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4" }, { "label": "Pinned served config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4/raw/183968f87ae4cedce3039313cac1fd43d112c578/config.json" }, { "label": "Pinned ModelOpt quantization config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4/raw/183968f87ae4cedce3039313cac1fd43d112c578/hf_quant_config.json" }, { "label": "Pinned BF16 config comparison", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-Base-BF16/raw/1493b9c9ca301d6b6e5b0bee90fef1990ba49ed4/config.json" }, { "label": "Nemotron-H runtime cache reference", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4/raw/4f0cf9daaeb7a4d5e23f80a00e7ed15f0e03caf6/modeling_nemotron_h.py" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4/resolve/183968f87ae4cedce3039313cac1fd43d112c578/model.safetensors.index.json" } ] }, { "id": "sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP", "name": "Qwen3.6-27B-Text-NVFP4-MTP NVFP4", "short_name": "Qwen3.6 27B Text NVFP4 MTP", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 338680, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "qwen3.6", "nvfp4", "quantized", "modelopt", "mtp", "speculative-decoding", "blackwell", "text-only", "text-generation", "conversational", "en", "zh", "ja", "ko", "fr", "de", "es", "it", "pt", "ru", "ar", "base_model:Qwen/Qwen3.6-27B", "base_model:quantized:Qwen/Qwen3.6-27B", "license:apache-2.0", "endpoints_compatible", "8-bit", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 27.320697856, "active_params_b": 25.624600064, "layers": 64, "attention_heads": 24, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "NVFP4", "weight_bytes_per_param": 0.5, "resident_weight_gb": 19.6374752, "active_weight_gb": 16.245279616, "kv_alloc_gb_per_1k_tokens": 0.032768, "kv_read_gb_per_1k_tokens": 0.032768, "kv_alloc_fixed_gb": 0.154927104, "kv_read_fixed_gb": 0.154927104, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, model card, pinned ModelOpt NVFP4 config and quantization config, current base text-config comparison, and direct range-read safetensors header. Bound weights separate the 19.637475200 GB resident NVFP4/BF16/F8/F32 text-only package from the 16.245279616 GB ordinary text-decode swept subset; production KV follows the model card's 256K FP8 KV launch. The audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP" }, { "label": "Pinned ModelOpt NVFP4 config", "url": "https://huggingface.co/sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP/raw/6f194695406a3bc88a00573187d5b2eecf984a99/config.json" }, { "label": "Pinned ModelOpt quantization config", "url": "https://huggingface.co/sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP/raw/6f194695406a3bc88a00573187d5b2eecf984a99/hf_quant_config.json" }, { "label": "Base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP/resolve/6f194695406a3bc88a00573187d5b2eecf984a99/model.safetensors" } ] }, { "id": "Qwen/Qwen2.5-VL-72B-Instruct", "name": "Qwen2.5-VL-72B-Instruct BF16", "short_name": "Qwen2.5 VL 72B", "license": "other", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 389971, "tags": [ "transformers", "safetensors", "qwen2_5_vl", "image-text-to-text", "multimodal", "conversational", "en", "arxiv:2309.00071", "arxiv:2409.12191", "arxiv:2308.12966", "license:other", "eval-results", "text-generation-inference", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2_5_vl (Qwen2_5_VLForConditionalGeneration)", "total_params_b": 73.410777344, "active_params_b": 71.46049536, "layers": 80, "attention_heads": 64, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 128000, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 146.821554688, "active_weight_gb": 142.92099072, "kv_alloc_gb_per_1k_tokens": 0.32768, "kv_read_gb_per_1k_tokens": 0.32768, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, pinned served config, model card, safetensors index, direct range-read safetensors shard headers, and Qwen2.5 72B text-profile comparison. Bound weights separate the 146.821554688 GB resident BF16 multimodal package from the 142.920990720 GB ordinary text/logit swept subset; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-72B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-VL-72B-Instruct" }, { "label": "Pinned served config", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-72B-Instruct/raw/89c86200743eec961a297729e7990e8f2ddbc4c5/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-72B-Instruct/resolve/89c86200743eec961a297729e7990e8f2ddbc4c5/model.safetensors.index.json" }, { "label": "Qwen2.5 72B text profile comparison", "url": "https://huggingface.co/Qwen/Qwen2.5-72B-Instruct/raw/495f39366efef23836d0cfae4fbe635880d2be31/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8", "name": "Qwen3-Coder-480B-A35B-Instruct-FP8 FP8", "short_name": "Qwen3 Coder 480B A35B Instruct FP8", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 531878, "tags": [ "transformers", "safetensors", "qwen3_moe", "text-generation", "conversational", "arxiv:2505.09388", "license:apache-2.0", "endpoints_compatible", "fp8", "deploy:azure", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_moe (Qwen3MoeForCausalLM)", "total_params_b": 480.184064, "active_params_b": 35, "layers": 62, "attention_heads": 96, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 160, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "FP8/BF16", "weight_bytes_per_param": 1, "resident_weight_gb": 482.141974528, "active_weight_gb": 35.54059264, "kv_alloc_gb_per_1k_tokens": 0.253952, "kv_read_gb_per_1k_tokens": 0.253952, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, model card, served FP8 config, BF16 config comparison, safetensors index, and direct safetensors header range reads. Active traffic is fixed decode traffic plus 8 routed expert groups; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8/raw/003f183a92fbe5b9a8325aaa8b2ae797c91dd90f/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8" }, { "label": "Safetensors index", "url": "https://huggingface.co/Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8/resolve/003f183a92fbe5b9a8325aaa8b2ae797c91dd90f/model.safetensors.index.json" }, { "label": "BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3-Coder-480B-A35B-Instruct/raw/9d90cf8fca1bf7b7acca42d3fc9ae694a2194069/config.json" } ] }, { "id": "cyankiwi/Qwen3.5-4B-AWQ-4bit", "name": "Qwen3.5-4B-AWQ-4bit 4-bit", "short_name": "Qwen3.5 4B AWQ 4bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 363086, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "conversational", "base_model:Qwen/Qwen3.5-4B", "base_model:quantized:Qwen/Qwen3.5-4B", "license:apache-2.0", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 4.659865088, "active_params_b": 4.205751296, "layers": 32, "attention_heads": 16, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "compressed-tensors AWQ INT4/BF16", "weight_bytes_per_param": 0.867045953412804, "resident_weight_gb": 4.040317168, "active_weight_gb": 3.286590592, "kv_alloc_gb_per_1k_tokens": 0.032768, "kv_read_gb_per_1k_tokens": 0.032768, "kv_alloc_fixed_gb": 0.051904512, "kv_read_fixed_gb": 0.051904512, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, pinned compressed-tensors config, Qwen3.5 4B base config comparison, model card, recipe, safetensors index, direct shard header range read, and the Transformers qwen3_5 implementation. Bound weights separate the exact 4.040317168 GB resident package from 3.286590592 GB ordinary text-decode swept traffic; the audited profile supersedes the generated ideal 4-bit estimate and adds the fixed DeltaNet runtime state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/cyankiwi/Qwen3.5-4B-AWQ-4bit" }, { "label": "Model config", "url": "https://huggingface.co/cyankiwi/Qwen3.5-4B-AWQ-4bit/raw/b878ff99819d1edba0d167876eaa2e4faef5d9d7/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-4B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3.5-4B-AWQ-4bit" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3.5-4B/raw/851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/cyankiwi/Qwen3.5-4B-AWQ-4bit/resolve/b878ff99819d1edba0d167876eaa2e4faef5d9d7/model.safetensors.index.json" } ] }, { "id": "stepfun-ai/Step3-VL-10B", "name": "Step3-VL-10B BF16", "short_name": "Step3 VL 10B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 351926, "tags": [ "safetensors", "step_robotics", "image-text-to-text", "conversational", "custom_code", "arxiv:2601.09668", "base_model:stepfun-ai/Step3-VL-10B-Base", "base_model:finetune:stepfun-ai/Step3-VL-10B-Base", "license:apache-2.0", "region:us" ], "architecture": { "type": "Dense multimodal transformer", "detail": "step_robotics (StepVLForConditionalGeneration + Qwen3 text decoder)", "total_params_b": 10.171750144, "active_params_b": 7.568405504, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 65536, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 20.343500288, "active_weight_gb": 15.136811008, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, model card, pinned config, base config comparison, remote modeling code, safetensors index, and direct shard header range reads. Bound weights separate the exact 20.343500288 GB resident BF16 multimodal package from 15.136811008 GB ordinary cached text-decode swept traffic; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/stepfun-ai/Step3-VL-10B" }, { "label": "Model config", "url": "https://huggingface.co/stepfun-ai/Step3-VL-10B/raw/5026053b0c2f5dfaa08fc2d149384162c3c8bca1/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/stepfun-ai/Step3-VL-10B-Base" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/stepfun-ai/Step3-VL-10B" }, { "label": "Base model config", "url": "https://huggingface.co/stepfun-ai/Step3-VL-10B-Base/raw/b84b55f58f5b65d108ef5993f4b71e11d421159c/config.json" }, { "label": "Remote modeling code", "url": "https://huggingface.co/stepfun-ai/Step3-VL-10B/raw/5026053b0c2f5dfaa08fc2d149384162c3c8bca1/modeling_step_vl.py" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/stepfun-ai/Step3-VL-10B/resolve/5026053b0c2f5dfaa08fc2d149384162c3c8bca1/model.safetensors.index.json" } ] }, { "id": "zai-org/chatglm2-6b", "name": "chatglm2-6b FP16", "short_name": "chatglm2 6b", "license": "unknown", "hf_pipeline_tag": "unknown", "hf_downloads": 413791, "tags": [ "transformers", "pytorch", "chatglm", "glm", "thudm", "custom_code", "zh", "en", "arxiv:2103.10360", "arxiv:2210.02414", "arxiv:1911.02150", "arxiv:2406.12793", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "chatglm (ChatGLMModel)", "total_params_b": 6.243584032, "active_params_b": 5.977245728, "layers": 28, "attention_heads": 32, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "FP16", "weight_bytes_per_param": 2, "resident_weight_gb": 12.487168064, "active_weight_gb": 11.954491456, "kv_alloc_gb_per_1k_tokens": 0.028672, "kv_read_gb_per_1k_tokens": 0.028672, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, model card, pinned config, remote modeling code, PyTorch shard index, and deterministic config tensor formulas. Bound weights separate the exact 12.487168064 GB resident FP16 PyTorch package from 11.954491456 GB ordinary text-decode swept traffic; the audited profile supersedes the generated metadata estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/zai-org/chatglm2-6b" }, { "label": "Model config", "url": "https://huggingface.co/zai-org/chatglm2-6b/raw/d2e2d91789248536a747d9ce60642a336444186c/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/zai-org/chatglm2-6b" }, { "label": "Remote modeling code", "url": "https://huggingface.co/zai-org/chatglm2-6b/raw/d2e2d91789248536a747d9ce60642a336444186c/modeling_chatglm.py" }, { "label": "PyTorch shard index", "url": "https://huggingface.co/zai-org/chatglm2-6b/resolve/d2e2d91789248536a747d9ce60642a336444186c/pytorch_model.bin.index.json" } ] }, { "id": "QuantTrio/gemma-4-31B-it-AWQ", "name": "gemma-4-31B-it-AWQ 4-bit", "short_name": "gemma 4 31B it", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 501734, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "vLLM", "AWQ", "conversational", "base_model:google/gemma-4-31B-it", "base_model:quantized:google/gemma-4-31B-it", "license:apache-2.0", "endpoints_compatible", "4-bit", "awq", "region:us" ], "architecture": { "type": "Dense multimodal transformer", "detail": "gemma4_text (Gemma4ForConditionalGeneration, AWQ)", "total_params_b": 31.273088876, "active_params_b": 30.69734534, "layers": 60, "attention_heads": 32, "kv_heads": 16, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "AWQ 4-bit", "weight_bytes_per_param": 0.6542301038801934, "resident_weight_gb": 20.459796184, "active_weight_gb": 19.308309112, "kv_alloc_gb_per_1k_tokens": 0.83968, "kv_read_gb_per_1k_tokens": 0.83968, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, pinned AWQ config, Google base config comparison, safetensors index, and direct shard header range reads. Bound weights separate the exact 20.459796184 GB resident AWQ package from 19.308309112 GB ordinary text-decode swept traffic; the audited profile supersedes the generated ideal 4-bit estimate. Legacy KV fields are the short-context unsaturated Gemma 4 layered-KV approximation." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/QuantTrio/gemma-4-31B-it-AWQ" }, { "label": "Model config", "url": "https://huggingface.co/QuantTrio/gemma-4-31B-it-AWQ/raw/200d06ee83eb49a03c6e3120dbf7b09191eb1539/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-31B-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/QuantTrio/gemma-4-31B-it-AWQ" }, { "label": "Base model config", "url": "https://huggingface.co/google/gemma-4-31B-it/raw/3548789868c5356dbf307c98e6f609007b82b3eb/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/QuantTrio/gemma-4-31B-it-AWQ/resolve/200d06ee83eb49a03c6e3120dbf7b09191eb1539/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen3-VL-32B-Thinking-FP8", "name": "Qwen3-VL-32B-Thinking-FP8 FP8", "short_name": "Qwen3 VL 32B Thinking FP8", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 306899, "tags": [ "transformers", "safetensors", "qwen3_vl", "image-text-to-text", "conversational", "arxiv:2505.09388", "arxiv:2502.13923", "arxiv:2409.12191", "arxiv:2308.12966", "license:apache-2.0", "endpoints_compatible", "fp8", "deploy:azure", "region:us" ], "architecture": { "type": "Dense multimodal transformer", "detail": "qwen3_vl_text (Qwen3VLForConditionalGeneration, FP8)", "total_params_b": 33.359294704, "active_params_b": 31.986115584, "layers": 64, "attention_heads": 64, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "FP8/BF16/F32", "weight_bytes_per_param": 1.0646740958747338, "resident_weight_gb": 35.516776928, "active_weight_gb": 32.770418688, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, model card, pinned FP8 config, BF16 Thinking config comparison, audited BF16 Instruct profile comparison, and direct safetensors shard header range reads. Bound weights separate the exact 35.516776928 GB resident FP8/BF16/F32 package from 32.770418688 GB ordinary text-decode swept traffic; the audited profile supersedes the generated FP8 estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-32B-Thinking-FP8" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-VL-32B-Thinking-FP8/raw/3eee143c9b355ba456390568e5970e2e60f5582b/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-VL-32B-Thinking-FP8" }, { "label": "BF16 Thinking base model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-32B-Thinking" }, { "label": "BF16 Thinking base config", "url": "https://huggingface.co/Qwen/Qwen3-VL-32B-Thinking/raw/7edd10ffd1196091948fb245ff63e406ccb2d4d1/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Qwen/Qwen3-VL-32B-Thinking-FP8/resolve/3eee143c9b355ba456390568e5970e2e60f5582b/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen3-4B-GGUF", "name": "Qwen3-4B-GGUF GGUF", "short_name": "Qwen3 4B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 388276, "tags": [ "gguf", "text-generation", "arxiv:2309.00071", "arxiv:2505.09388", "base_model:Qwen/Qwen3-4B", "base_model:quantized:Qwen/Qwen3-4B", "license:apache-2.0", "endpoints_compatible", "conversational", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM, selected Q4_K_M GGUF)", "total_params_b": 4.022468096, "active_params_b": 4.022468096, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q4_K_M", "weight_bytes_per_param": 0.6208328310877919, "resident_weight_gb": 2.497280256, "active_weight_gb": 2.491323904, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, model card, pinned Qwen3 4B config, existing BF16 profile comparison, selected linked-object HEAD check, and direct GGUF header range read. The authoritative bounds profile uses the API-selected Qwen3-4B-Q4_K_M.gguf artifact: 2.497280256 GB resident, 2.491323904 GB swept text traffic, 0.005956352 GB resident-only GGUF metadata/header/file overhead, and full-context Qwen3 FP16 KV." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-4B-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3-4B/raw/1cfa9a7208912126459214e8b04321603b3df60c/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-4B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-4B-GGUF" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/Qwen/Qwen3-4B-GGUF/resolve/bc640142c66e1fdd12af0bd68f40445458f3869b/Qwen3-4B-Q4_K_M.gguf" } ] }, { "id": "google/gemma-2-2b-it", "name": "gemma-2-2b-it BF16", "short_name": "gemma 2 2b it", "license": "gemma", "hf_pipeline_tag": "text-generation", "hf_downloads": 445661, "tags": [ "transformers", "safetensors", "gemma2", "text-generation", "conversational", "arxiv:2009.03300", "arxiv:1905.07830", "arxiv:1911.11641", "arxiv:1904.09728", "arxiv:1905.10044", "arxiv:1907.10641", "arxiv:1811.00937", "arxiv:1809.02789", "arxiv:1911.01547", "arxiv:1705.03551", "arxiv:2107.03374", "arxiv:2108.07732", "arxiv:2110.14168", "arxiv:2009.11462", "arxiv:2101.11718", "arxiv:2110.08193", "arxiv:1804.09301", "arxiv:2109.07958", "arxiv:1804.06876", "arxiv:2103.03874", "arxiv:2304.06364", "arxiv:1903.00161", "arxiv:2206.04615", "arxiv:2203.09509", "arxiv:2403.13793", "base_model:google/gemma-2-2b", "base_model:finetune:google/gemma-2-2b", "license:gemma", "eval-results", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma2 (gated config unavailable)", "total_params_b": 2.614341888, "active_params_b": 2.614341888, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 5.228683776, "active_weight_gb": 5.228683776, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manual fail-closed correction on 2026-07-05 from current Hugging Face API metadata and direct hf download access checks. The repo and base repo are manually gated; config and tensor headers are inaccessible, so production bounds use the unsupported profile and must not infer KV geometry or swept traffic from this legacy row." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-2-2b-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-2-2b-it" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-2-2b" }, { "label": "Base Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-2-2b" } ] }, { "id": "RedHatAI/diffusiongemma-26B-A4B-it-NVFP4", "name": "diffusiongemma-26B-A4B-it-NVFP4 NVFP4", "short_name": "diffusiongemma 26B A4B it NVFP4", "license": "unknown", "hf_pipeline_tag": "unknown", "hf_downloads": 539635, "tags": [ "safetensors", "diffusion_gemma", "nvfp4", "vllm", "compressed-tensors", "llm-compressor", "base_model:google/diffusiongemma-26B-A4B-it", "base_model:quantized:google/diffusiongemma-26B-A4B-it", "8-bit", "region:us" ], "architecture": { "type": "Block diffusion MoE transformer", "detail": "diffusion_gemma (DiffusionGemmaForBlockDiffusion)", "total_params_b": 25.823781228, "active_params_b": 4, "layers": 30, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "NVFP4", "weight_bytes_per_param": 0.5, "resident_weight_gb": 18.048607656, "active_weight_gb": 4.859463564, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manual fail-closed correction on 2026-07-05 from current HF API commit 8484b309d29a474b098ad57c47a9045e1fc4b48a, pinned config, recipe, generation config, model card, safetensors index, and direct range-read header. The exact compressed-tensors resident package is 18.048607656 GB. Production bounds are disabled by the unsupported profile because DiffusionGemma uses block diffusion over a 256-token canvas, which Bounds Engine v1 does not model as ordinary autoregressive decode." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/RedHatAI/diffusiongemma-26B-A4B-it-NVFP4" }, { "label": "Model config", "url": "https://huggingface.co/RedHatAI/diffusiongemma-26B-A4B-it-NVFP4/raw/8484b309d29a474b098ad57c47a9045e1fc4b48a/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/diffusiongemma-26B-A4B-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/RedHatAI/diffusiongemma-26B-A4B-it-NVFP4" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/RedHatAI/diffusiongemma-26B-A4B-it-NVFP4/raw/8484b309d29a474b098ad57c47a9045e1fc4b48a/model.safetensors.index.json" } ] }, { "id": "ggml-org/gemma-4-26B-A4B-it-GGUF", "name": "gemma-4-26B-A4B-it-GGUF BF16 GGUF", "short_name": "gemma 4 26B A4B it", "license": "apache-2.0", "hf_pipeline_tag": "unknown", "hf_downloads": 433222, "tags": [ "gguf", "base_model:google/gemma-4-26B-A4B-it", "base_model:quantized:google/gemma-4-26B-A4B-it", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "MoE transformer", "detail": "gemma4_text", "total_params_b": 25.233142046, "active_params_b": 3.8, "layers": 30, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "GGUF BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 50.505134624, "active_weight_gb": 7.668089976, "kv_alloc_gb_per_1k_tokens": 0.20992, "kv_read_gb_per_1k_tokens": 0.20992, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, README, base profile comparison, linked GGUF file sizes, and a direct GGUF tensor-index range read. The authoritative bounds profile uses the API-selected gemma-4-26B-A4B-it-bf16.gguf artifact: 50.505134624 GB resident, 4.813341816 GB fixed traffic, 0.356843520 GB per routed expert, 8 routed experts per token, 128 routed experts, 1 shared expert charged in fixed traffic, and layered Gemma 4 hybrid KV. Legacy per-1k KV fields show the short-context unsaturated coefficient; long-context calculation must use the profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/ggml-org/gemma-4-26B-A4B-it-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/ggml-org/gemma-4-26B-A4B-it-GGUF" }, { "label": "Selected BF16 GGUF artifact", "url": "https://huggingface.co/ggml-org/gemma-4-26B-A4B-it-GGUF/resolve/ae4d537a6345467d1c86bb5cc0d4505ff3ebe0f3/gemma-4-26B-A4B-it-bf16.gguf" } ] }, { "id": "zai-org/GLM-4.5-Air", "name": "GLM-4.5-Air BF16/F32", "short_name": "GLM 4.5 Air", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 398671, "tags": [ "transformers", "safetensors", "glm4_moe", "text-generation", "conversational", "en", "zh", "arxiv:2508.06471", "license:mit", "eval-results", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "glm4_moe (Glm4MoeForCausalLM)", "total_params_b": 110.468824832, "active_params_b": 12.803378432, "layers": 46, "attention_heads": 96, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "BF16/F32", "weight_bytes_per_param": 2.000000106600211, "resident_weight_gb": 220.93766144, "active_weight_gb": 25.606756864, "kv_alloc_gb_per_1k_tokens": 0.188416, "kv_read_gb_per_1k_tokens": 0.188416, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, pinned served config, generation config, safetensors index, direct range-read shard headers, and the Transformers glm4_moe implementation. Bound weights separate the exact 220.937661440 GB resident BF16/F32 package from 25.606756864 GB ordinary text-decode swept traffic: 13.149673984 GB fixed traffic plus 8 routed expert groups at 1.557135360 GB each. The input embedding and model.layers.46 next-token-prediction tensor package are resident-only for ordinary decode." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/zai-org/GLM-4.5-Air" }, { "label": "Model config", "url": "https://huggingface.co/zai-org/GLM-4.5-Air/raw/a24ceef6ce4f3536971efe9b778bdaa1bab18daa/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/zai-org/GLM-4.5-Air" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/zai-org/GLM-4.5-Air/raw/a24ceef6ce4f3536971efe9b778bdaa1bab18daa/model.safetensors.index.json" } ] }, { "id": "deepseek-ai/DeepSeek-R1-Distill-Llama-70B", "name": "DeepSeek-R1-Distill-Llama-70B BF16", "short_name": "DeepSeek R1 Distill Llama 70B", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 571006, "tags": [ "transformers", "safetensors", "llama", "text-generation", "conversational", "arxiv:2501.12948", "license:mit", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "llama (LlamaForCausalLM)", "total_params_b": 70.553706496, "active_params_b": 69.503033344, "layers": 80, "attention_heads": 64, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 141.107412992, "active_weight_gb": 139.006066688, "kv_alloc_gb_per_1k_tokens": 0.32768, "kv_read_gb_per_1k_tokens": 0.32768, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, pinned served config, generation config, model card, safetensors index, direct range-read shard headers, and the Transformers Llama implementation. Bound weights separate the exact 141.107412992 GB resident BF16 package from 139.006066688 GB ordinary text-decode swept traffic; model.embed_tokens.weight is resident-only and lm_head.weight is stored separately and swept." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B" }, { "label": "Model config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B/raw/b1c0b44b4369b597ad119a196caf79a9c40e141e/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-R1-Distill-Llama-70B" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B/raw/b1c0b44b4369b597ad119a196caf79a9c40e141e/model.safetensors.index.json" } ] }, { "id": "bartowski/Llama-3.2-1B-Instruct-GGUF", "name": "Llama-3.2-1B-Instruct-GGUF GGUF", "short_name": "Llama 3.2 1B Instruct", "license": "llama3.2", "hf_pipeline_tag": "text-generation", "hf_downloads": 416350, "tags": [ "gguf", "facebook", "meta", "llama", "llama-3", "text-generation", "en", "de", "fr", "it", "pt", "hi", "es", "th", "base_model:meta-llama/Llama-3.2-1B-Instruct", "base_model:quantized:meta-llama/Llama-3.2-1B-Instruct", "license:llama3.2", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "llama", "total_params_b": 1.235814432, "active_params_b": 1.235814432, "layers": 16, "attention_heads": 32, "kv_heads": 8, "head_dim": 64, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF F16", "weight_bytes_per_param": 2.006446352942413, "resident_weight_gb": 2.47959536, "active_weight_gb": 2.471764096, "kv_alloc_gb_per_1k_tokens": 0.032768, "kv_read_gb_per_1k_tokens": 0.032768, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, model card, base API metadata, gated base-config access check, linked GGUF object HEAD checks, and a direct range-read GGUF tensor index. The authoritative bounds profile targets Llama-3.2-1B-Instruct-f16.gguf because API gguf.totalFileSize exactly matches that linked object: 2.479595360 GB resident, 2.471764096 GB swept tensor spans, and 0.007831264 GB resident-only GGUF metadata/header/file overhead." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/bartowski/Llama-3.2-1B-Instruct-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/bartowski/Llama-3.2-1B-Instruct-GGUF" }, { "label": "Selected F16 GGUF artifact", "url": "https://huggingface.co/bartowski/Llama-3.2-1B-Instruct-GGUF/resolve/067b946cf014b7c697f3654f621d577a3e3afd1c/Llama-3.2-1B-Instruct-f16.gguf" } ] }, { "id": "google/gemma-4-12B", "name": "gemma-4-12B BF16", "short_name": "gemma 4 12B", "license": "apache-2.0", "hf_pipeline_tag": "any-to-any", "hf_downloads": 394761, "tags": [ "transformers", "safetensors", "gemma4_unified", "image-text-to-text", "any-to-any", "license:apache-2.0", "eval-results", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_unified_text", "total_params_b": 11.959730224, "active_params_b": 11.90735032, "layers": 48, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 23.919460448, "active_weight_gb": 23.81470064, "kv_alloc_gb_per_1k_tokens": 0.335872, "kv_read_gb_per_1k_tokens": 0.335872, "notes": "Manually corrected on 2026-07-05 from current Hugging Face API metadata, model card, pinned served config, instruction-tuned config comparison, and direct safetensors header range read. The authoritative bounds profile records 23.919460448 GB resident BF16 tensors, 23.814700640 GB swept language-model traffic, and 0.104759808 GB resident-only multimodal projection tensors. Legacy per-1k KV fields show the short-context unsaturated coefficient; long-context calculation must use the layered audited profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-4-12B" }, { "label": "Model config", "url": "https://huggingface.co/google/gemma-4-12B/raw/56820d7d8cbe8e47975a53325439ed272e91cff2/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-12B" }, { "label": "Safetensors header source", "url": "https://huggingface.co/google/gemma-4-12B/resolve/56820d7d8cbe8e47975a53325439ed272e91cff2/model.safetensors" } ] }, { "id": "Qwen/Qwen3-4B-AWQ", "name": "Qwen3-4B-AWQ 4-bit", "short_name": "Qwen3 4B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 508651, "tags": [ "transformers", "safetensors", "qwen3", "text-generation", "conversational", "arxiv:2309.00071", "arxiv:2505.09388", "base_model:Qwen/Qwen3-4B", "base_model:quantized:Qwen/Qwen3-4B", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "4-bit", "awq", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 4.022468096, "active_params_b": 4.022468096, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "4-bit", "weight_bytes_per_param": 0.6627586766072886, "resident_weight_gb": 2.665925632, "active_weight_gb": 2.665925632, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Manually corrected on 2026-07-05 from current Hugging Face API metadata, model card, pinned served AWQ config, Qwen/Qwen3-4B base config comparison, and direct safetensors header range read. The authoritative bounds profile records 2.665925632 GB resident and swept tensor bytes: 1.816657920 GB I32 qweight, 0.014192640 GB I32 qzeros, 0.056770560 GB F16 scales, and 0.778304512 GB BF16 tensors. The file has no lm_head.weight, so the tied embedding remains swept." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-4B-AWQ" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-4B-AWQ/raw/74d4bd2bd4bff9cafc9345221320bffb08b406a3/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-4B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-4B-AWQ" }, { "label": "Safetensors header source", "url": "https://huggingface.co/Qwen/Qwen3-4B-AWQ/resolve/74d4bd2bd4bff9cafc9345221320bffb08b406a3/model.safetensors" } ] }, { "id": "cyankiwi/Qwen3.5-122B-A10B-AWQ-4bit", "name": "Qwen3.5-122B-A10B-AWQ-4bit 4-bit", "short_name": "Qwen3.5 122B A10B AWQ 4bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 349099, "tags": [ "transformers", "safetensors", "qwen3_5_moe", "image-text-to-text", "conversational", "base_model:Qwen/Qwen3.5-122B-A10B", "base_model:quantized:Qwen/Qwen3.5-122B-A10B", "license:apache-2.0", "endpoints_compatible", "compressed-tensors", "deploy:azure", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_5_moe_text", "total_params_b": 125.086497008, "active_params_b": 10, "layers": 48, "attention_heads": 32, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "compressed-tensors AWQ Int4 packed I32 + BF16/I64 side tensors", "weight_bytes_per_param": 0.6395757827871972, "resident_weight_gb": 80.00229424, "active_weight_gb": 12.807591936, "kv_alloc_gb_per_1k_tokens": 0.024576, "kv_read_gb_per_1k_tokens": 0.024576, "kv_alloc_fixed_gb": 0.154533888, "kv_read_fixed_gb": 0.154533888, "expert_param_b": 0.3983988744689234, "fixed_param_b": 16.8379448531796, "shared_expert_param_b": 0, "notes": "Manually corrected on 2026-07-05 from current Hugging Face API metadata, pinned compressed-tensors config, BF16 base config comparison, recipe.yaml, safetensors index, and direct range-read safetensors shard headers. Legacy fallback values mirror the audited profile: exact stored resident bytes, fixed always-on hybrid/shared-expert traffic, expected distinct routed experts, BF16 attention KV, and fixed BF16-convolution/F32-SSM DeltaNet recurrent state. The authoritative profile supersedes the generated 62.5433 GB resident estimate, the missing shared-expert count, and the all-layers full-context KV estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/cyankiwi/Qwen3.5-122B-A10B-AWQ-4bit" }, { "label": "Model config", "url": "https://huggingface.co/cyankiwi/Qwen3.5-122B-A10B-AWQ-4bit/raw/2e7164b035cdf708f05a1c5ce8f532ec39096d40/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B/raw/dc4d348443bc740c68e2d77492492c11606384d5/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3.5-122B-A10B-AWQ-4bit" }, { "label": "Safetensors index", "url": "https://huggingface.co/cyankiwi/Qwen3.5-122B-A10B-AWQ-4bit/resolve/2e7164b035cdf708f05a1c5ce8f532ec39096d40/model.safetensors.index.json" } ] }, { "id": "unsloth/Qwen2.5-VL-7B-Instruct-GGUF", "name": "Qwen2.5-VL-7B-Instruct-GGUF GGUF", "short_name": "Qwen2.5 VL 7B Instruct", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 390164, "tags": [ "transformers", "gguf", "qwen2_5_vl", "image-text-to-text", "multimodal", "unsloth", "en", "arxiv:2309.00071", "arxiv:2409.12191", "arxiv:2308.12966", "base_model:Qwen/Qwen2.5-VL-7B-Instruct", "base_model:quantized:Qwen/Qwen2.5-VL-7B-Instruct", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen2_5_vl (Qwen2_5_VLForConditionalGeneration)", "total_params_b": 7.615616512, "active_params_b": 7.070619136, "layers": 28, "attention_heads": 28, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 128000, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF BF16/F32 main file", "weight_bytes_per_param": 2.000869102585401, "resident_weight_gb": 15.237851776, "active_weight_gb": 14.141904896, "kv_alloc_gb_per_1k_tokens": 0.057344, "kv_read_gb_per_1k_tokens": 0.057344, "notes": "Manually corrected on 2026-07-05 from current Hugging Face API metadata, pinned Unsloth config, Qwen base config comparison, linked GGUF file sizes, and a direct GGUF tensor-index range read. The authoritative profile targets the API-selected BF16 main GGUF artifact: 15.237851776 GB resident file size, 14.141904896 GB swept text traffic, 1.095946880 GB resident-only token embedding plus GGUF overhead, and separate mmproj sidecars excluded unless a workload explicitly loads them." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/Qwen2.5-VL-7B-Instruct-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/unsloth/Qwen2.5-VL-7B-Instruct-GGUF/raw/68bb8bc4b7df5289c143aaec0ab477a7d4051aab/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct/raw/cc594898137f460bfe9f0759e9844b3ce807cfb5/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen2.5-VL-7B-Instruct-GGUF" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/unsloth/Qwen2.5-VL-7B-Instruct-GGUF/resolve/68bb8bc4b7df5289c143aaec0ab477a7d4051aab/Qwen2.5-VL-7B-Instruct-BF16.gguf" } ] }, { "id": "nvidia/Nemotron-Mini-4B-Instruct", "name": "Nemotron-Mini-4B-Instruct BF16", "short_name": "Nemotron Mini 4B", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 390524, "tags": [ "transformers", "pytorch", "nemo", "nvidia", "llama-3", "text-generation", "en", "arxiv:2402.16819", "arxiv:2407.14679", "license:other", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "nemotron (NemotronForCausalLM)", "total_params_b": 4.190509056, "active_params_b": 3.404077056, "layers": 32, "attention_heads": 24, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 4096, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 8.381018112, "active_weight_gb": 6.808154112, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "notes": "Manually corrected on 2026-07-05 from current Hugging Face API metadata, model card, served config, Minitron base config comparison, linked artifact HEAD checks, and direct PyTorch zip/pickle metadata range reads. The authoritative profile records 4.190509056B BF16 tensor params / 8.381018112 GB resident tensor payload and 3.404077056B params / 6.808154112 GB ordinary text-decode swept traffic after excluding only the resident input embedding." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/Nemotron-Mini-4B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/nvidia/Nemotron-Mini-4B-Instruct/raw/791833e92ebddb0bc2c1007f6d2b6764f886a2ae/config.json" }, { "label": "PyTorch checkpoint artifact", "url": "https://huggingface.co/nvidia/Nemotron-Mini-4B-Instruct/resolve/791833e92ebddb0bc2c1007f6d2b6764f886a2ae/pytorch_model.bin" } ] }, { "id": "google/gemma-3n-E2B-it", "name": "gemma-3n-E2B-it BF16", "short_name": "gemma 3n E2B it", "license": "gemma", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 409872, "tags": [ "transformers", "safetensors", "gemma3n", "image-text-to-text", "automatic-speech-recognition", "automatic-speech-translation", "audio-text-to-text", "video-text-to-text", "conversational", "arxiv:1905.07830", "arxiv:1905.10044", "arxiv:1911.11641", "arxiv:1904.09728", "arxiv:1705.03551", "arxiv:1911.01547", "arxiv:1907.10641", "arxiv:1903.00161", "arxiv:2210.03057", "arxiv:2502.12404", "arxiv:2411.19799", "arxiv:2009.03300", "arxiv:2502.21228", "arxiv:2311.12022", "arxiv:2403.07974", "arxiv:2108.07732", "arxiv:2107.03374", "base_model:google/gemma-3n-E4B-it", "base_model:finetune:google/gemma-3n-E4B-it", "license:gemma", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "unknown", "total_params_b": 5.439438272, "active_params_b": 5.439438272, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 10.878876544, "active_weight_gb": 10.878876544, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manually reviewed on 2026-07-05 from current Hugging Face API metadata and gated access checks. The API records 5.439438272B BF16 safetensors parameters / 10.878876544 GB, but config and tensor headers are inaccessible in this audit environment, so the profile deliberately fails closed instead of guessing KV geometry, context, or swept traffic." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-3n-E2B-it" }, { "label": "Gated config access check", "url": "https://huggingface.co/google/gemma-3n-E2B-it/raw/5e092ebca197cdcd8d8b195040accf22693501bc/config.json" } ] }, { "id": "kaitchup/Phi-3-mini-4k-instruct-gptq-4bit", "name": "Phi-3-mini-4k-instruct-gptq-4bit 4-bit", "short_name": "Phi 3 mini 4k instruct gptq 4bit", "license": "unknown", "hf_pipeline_tag": "text-generation", "hf_downloads": 319947, "tags": [ "transformers", "safetensors", "phi3", "text-generation", "conversational", "custom_code", "arxiv:1910.09700", "text-generation-inference", "endpoints_compatible", "4-bit", "gptq", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "phi3 (Phi3ForCausalLM)", "total_params_b": 3.82209536, "active_params_b": 3.723594752, "layers": 32, "attention_heads": 32, "kv_heads": 32, "head_dim": 96, "max_context_tokens": 4096, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "4-bit", "weight_bytes_per_param": 0.596892443834787, "resident_weight_gb": 2.28137984, "active_weight_gb": 2.084378624, "kv_alloc_gb_per_1k_tokens": 0.393216, "kv_read_gb_per_1k_tokens": 0.393216, "notes": "Manually corrected on 2026-07-05 from current Hugging Face API metadata, served GPTQ config, pinned Microsoft base config comparison, linked artifact HEAD check, and direct safetensors header range read. The authoritative profile records 2.281379840 GB resident GPTQ tensor payload, 2.084378624 GB ordinary text-decode swept traffic, and a 2048-token FP16 sliding-window KV cache." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/kaitchup/Phi-3-mini-4k-instruct-gptq-4bit" }, { "label": "Model config", "url": "https://huggingface.co/kaitchup/Phi-3-mini-4k-instruct-gptq-4bit/raw/8d0c383a88f9b355bbc7ac0c36625fa7b9421e71/config.json" }, { "label": "Safetensors artifact", "url": "https://huggingface.co/kaitchup/Phi-3-mini-4k-instruct-gptq-4bit/resolve/8d0c383a88f9b355bbc7ac0c36625fa7b9421e71/model.safetensors" } ] }, { "id": "mlx-community/gpt-oss-20b-MXFP4-Q8", "name": "gpt-oss-20b-MXFP4-Q8 MXFP4", "short_name": "gpt oss 20b MXFP4 Q8", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 371761, "tags": [ "mlx", "safetensors", "gpt_oss", "vllm", "text-generation", "conversational", "base_model:openai/gpt-oss-20b", "base_model:quantized:openai/gpt-oss-20b", "license:apache-2.0", "4-bit", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "gpt_oss (GptOssForCausalLM)", "total_params_b": 20.914755648, "active_params_b": 3.6, "layers": 24, "attention_heads": 64, "kv_heads": 8, "head_dim": 64, "max_context_tokens": 131072, "routed_experts": 32, "routed_experts_per_token": 4, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "MLX MXFP4/Q8", "weight_bytes_per_param": 0.5773970956794225, "resident_weight_gb": 12.076119168, "active_weight_gb": 2.565875328, "kv_alloc_gb_per_1k_tokens": 0.02460745728, "kv_read_gb_per_1k_tokens": 0.024674304, "notes": "Manually corrected on 2026-07-05 from current Hugging Face API metadata, MLX model card, served config, pinned openai/gpt-oss-20b config comparison, safetensors index metadata, linked shard HEAD checks, and direct safetensors header range reads. The authoritative profile records 12.076119168 GB resident MLX quantized tensor payload, 2.565875328 GB single-token ordinary text-decode traffic, and layered BF16 KV with 12 full-context layers plus 12 sliding-window layers." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/mlx-community/gpt-oss-20b-MXFP4-Q8" }, { "label": "Model config", "url": "https://huggingface.co/mlx-community/gpt-oss-20b-MXFP4-Q8/raw/773a7da77e569019bb0fd17a554b263738d669a3/config.json" }, { "label": "Safetensors index", "url": "https://huggingface.co/mlx-community/gpt-oss-20b-MXFP4-Q8/raw/773a7da77e569019bb0fd17a554b263738d669a3/model.safetensors.index.json" }, { "label": "Base model card", "url": "https://huggingface.co/openai/gpt-oss-20b" } ] }, { "id": "lmstudio-community/Qwen3.5-9B-GGUF", "name": "Qwen3.5-9B-GGUF GGUF", "short_name": "Qwen3.5 9B", "license": "apache-2.0", "hf_pipeline_tag": "unknown", "hf_downloads": 391812, "tags": [ "gguf", "base_model:Qwen/Qwen3.5-9B", "base_model:quantized:Qwen/Qwen3.5-9B", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense multimodal transformer", "detail": "qwen3_5_text (selected Q4_K_M GGUF)", "total_params_b": 8.953803264, "active_params_b": 8.953803264, "layers": 32, "attention_heads": 16, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q4_K_M", "weight_bytes_per_param": 0.6284529702170593, "resident_weight_gb": 5.627044256, "active_weight_gb": 5.04394752, "kv_alloc_gb_per_1k_tokens": 0.032768, "kv_read_gb_per_1k_tokens": 0.032768, "kv_alloc_fixed_gb": 0.051904512, "kv_read_fixed_gb": 0.051904512, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, model card, pinned Qwen3.5 9B config, linked GGUF object HEAD checks, direct GGUF header range read, and the Transformers qwen3_5 implementation. The authoritative bounds profile uses the API-selected Qwen3.5-9B-Q4_K_M.gguf artifact: 5.627044256 GB resident, 5.04394752 GB swept text traffic, 0.583096736 GB resident-only token embedding plus GGUF metadata/header/file overhead, and Qwen3.5 hybrid KV plus fixed DeltaNet state. The separate BF16 mmproj sidecar is not included unless explicitly loaded." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/Qwen3.5-9B-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3.5-9B/raw/c202236235762e1c871ad0ccb60c8ee5ba337b9a/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-9B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3.5-9B-GGUF" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/lmstudio-community/Qwen3.5-9B-GGUF/resolve/1379f25c6b505a3fc737bd7818cb09389cf807c1/Qwen3.5-9B-Q4_K_M.gguf" } ] }, { "id": "deepseek-ai/deepseek-coder-6.7b-instruct", "name": "deepseek-coder-6.7b-instruct BF16", "short_name": "deepseek coder 6.7b", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 382398, "tags": [ "transformers", "pytorch", "safetensors", "llama", "text-generation", "conversational", "license:other", "text-generation-inference", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "llama (LlamaForCausalLM)", "total_params_b": 6.740512768, "active_params_b": 6.608392192, "layers": 32, "attention_heads": 32, "kv_heads": 32, "head_dim": 128, "max_context_tokens": 16384, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 13.481025536, "active_weight_gb": 13.216784384, "kv_alloc_gb_per_1k_tokens": 0.524288, "kv_read_gb_per_1k_tokens": 0.524288, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, model card, pinned served config, base config comparison, safetensors index, and direct shard header range reads. Bound weights separate the exact 13.481025536 GB resident BF16 package from 13.216784384 GB ordinary text-decode swept traffic; the audited profile supersedes the generated rounded dense estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-instruct" }, { "label": "Model config", "url": "https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-instruct/raw/e5d64addd26a6a1db0f9b863abf6ee3141936807/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/deepseek-coder-6.7b-instruct" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-instruct/resolve/e5d64addd26a6a1db0f9b863abf6ee3141936807/model.safetensors.index.json" } ] }, { "id": "ggml-org/gemma-4-12B-it-GGUF", "name": "gemma-4-12B-it-GGUF GGUF", "short_name": "gemma 4 12B it", "license": "apache-2.0", "hf_pipeline_tag": "unknown", "hf_downloads": 286749, "tags": [ "gguf", "base_model:google/gemma-4-12B-it", "base_model:quantized:google/gemma-4-12B-it", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_unified_text", "total_params_b": 11.907350576, "active_params_b": 11.907350576, "layers": 48, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16 GGUF", "weight_bytes_per_param": 2, "resident_weight_gb": 23.832064928, "active_weight_gb": 23.816242688, "kv_alloc_gb_per_1k_tokens": 0.335872, "kv_read_gb_per_1k_tokens": 0.335872, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, README, Google base config/profile, linked GGUF file sizes, and a GGUF tensor-index range read. The authoritative bounds profile uses the selected gemma-4-12B-it-bf16.gguf artifact: 23.832064928 GB resident, 23.816242688 GB swept text traffic, 0.015822240 GB resident-only GGUF metadata/header bytes, and layered Gemma 4 hybrid KV. Separate mmproj GGUF files are not included unless explicitly loaded. Legacy per-1k KV fields show the short-context unsaturated coefficient; long-context calculation must use the profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/ggml-org/gemma-4-12B-it-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/google/gemma-4-12B-it/raw/5926caa4ec0cac5cbfadaf4077420520de1d5205/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-12B-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/ggml-org/gemma-4-12B-it-GGUF" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/ggml-org/gemma-4-12B-it-GGUF/resolve/44ee90c4b61e888ac5b318a54ec7a94df61e9cd7/gemma-4-12B-it-bf16.gguf" } ] }, { "id": "google/medgemma-1.5-4b-it", "name": "medgemma-1.5-4b-it BF16", "short_name": "medgemma 1.5 4b it", "license": "other", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 310950, "tags": [ "transformers", "safetensors", "gemma3", "image-text-to-text", "medical", "radiology", "clinical-reasoning", "dermatology", "pathology", "ophthalmology", "chest-x-ray", "conversational", "arxiv:2303.15343", "arxiv:2604.05081", "arxiv:2406.19578", "arxiv:2405.03162", "arxiv:2106.14463", "arxiv:2009.13081", "arxiv:2203.14371", "arxiv:1909.06146", "arxiv:2102.09542", "arxiv:2411.15640", "arxiv:2404.05590", "arxiv:2501.18362", "arxiv:1605.01397", "arxiv:1901.07031", "arxiv:2403.17834", "arxiv:2402.16040", "license:other", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense multimodal transformer", "detail": "gemma3 (gated config inaccessible)", "total_params_b": 4.300079472, "active_params_b": 4.300079472, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 8.600158944, "active_weight_gb": 8.600158944, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manually corrected on 2026-07-05 from current Hugging Face API metadata. The repo is auto-gated; raw config, README, generation config, safetensors index, safetensors shards, ranged safetensors bytes, and hf download config.json all returned access denied. Production bounds use the unsupported profile and must not infer KV geometry from this row." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/medgemma-1.5-4b-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/medgemma-1.5-4b-it" } ] }, { "id": "groxaxo/Qwen3.6-27B-GPTQ-Pro-4bit", "name": "Qwen3.6-27B-GPTQ-Pro-4bit 4-bit", "short_name": "Qwen3.6 27B GPTQ Pro 4bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 602309, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "text-generation-inference", "qwen", "gptq", "marlin", "foem", "conversational", "en", "base_model:Qwen/Qwen3.6-27B", "base_model:quantized:Qwen/Qwen3.6-27B", "license:apache-2.0", "endpoints_compatible", "4-bit", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 27.35672856, "active_params_b": 25.624600064, "layers": 64, "attention_heads": 24, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GPTQ 4-bit", "weight_bytes_per_param": 0.6839176018786363, "resident_weight_gb": 18.709748192, "active_weight_gb": 15.2454912, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, README, pinned GPTQ config and quantize_config, Qwen3.6 27B base config comparison, safetensors index, direct shard header range reads, and the Transformers qwen3_5 implementation. Bound weights separate the exact 18.709748192 GB resident package from 15.245491200 GB ordinary text-decode swept traffic; the audited profile supersedes the generated ideal 4-bit estimate and adds the fixed DeltaNet runtime state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/groxaxo/Qwen3.6-27B-GPTQ-Pro-4bit" }, { "label": "Model config", "url": "https://huggingface.co/groxaxo/Qwen3.6-27B-GPTQ-Pro-4bit/raw/62b96ffbca486698719ec67f7f414251ad2347ea/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.6-27B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/groxaxo/Qwen3.6-27B-GPTQ-Pro-4bit" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/groxaxo/Qwen3.6-27B-GPTQ-Pro-4bit/resolve/62b96ffbca486698719ec67f7f414251ad2347ea/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen3-VL-32B-Instruct-FP8", "name": "Qwen3-VL-32B-Instruct-FP8 FP8", "short_name": "Qwen3 VL 32B Instruct FP8", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 508282, "tags": [ "transformers", "safetensors", "qwen3_vl", "image-text-to-text", "conversational", "arxiv:2505.09388", "arxiv:2502.13923", "arxiv:2409.12191", "arxiv:2308.12966", "base_model:Qwen/Qwen3-VL-32B-Instruct", "base_model:quantized:Qwen/Qwen3-VL-32B-Instruct", "license:apache-2.0", "endpoints_compatible", "fp8", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_vl_text", "total_params_b": 33.359294704, "active_params_b": 31.986115584, "layers": 64, "attention_heads": 64, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "FP8", "weight_bytes_per_param": 1.0646740958747338, "resident_weight_gb": 35.516776928, "active_weight_gb": 32.770418688, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, README, pinned FP8 config, current BF16 Instruct config comparison, existing BF16 Instruct profile, safetensors index, and direct shard header range reads. Bound weights separate the exact 35.516776928 GB resident FP8/BF16/F32 package from 32.770418688 GB ordinary text-decode swept traffic; BF16 full-context KV is retained because no FP8 KV cache scheme is declared." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-32B-Instruct-FP8" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-VL-32B-Instruct-FP8/raw/4bf2c2f39c37c0fede78bede4056e1f18cdf8109/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-32B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-VL-32B-Instruct-FP8" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Qwen/Qwen3-VL-32B-Instruct-FP8/resolve/4bf2c2f39c37c0fede78bede4056e1f18cdf8109/model.safetensors.index.json" } ] }, { "id": "casperhansen/deepseek-r1-distill-qwen-32b-awq", "name": "deepseek-r1-distill-qwen-32b-awq 4-bit", "short_name": "deepseek r1 distill qwen 32b", "license": "mit", "hf_pipeline_tag": "unknown", "hf_downloads": 272214, "tags": [ "safetensors", "qwen2", "license:mit", "4-bit", "awq", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 32.763876352, "active_params_b": 31.985308672, "layers": 64, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "AWQ 4-bit", "weight_bytes_per_param": 0.5899425530831646, "resident_weight_gb": 19.328804864, "active_weight_gb": 17.771669504, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, pinned served AWQ config, minimal README, audited BF16 DeepSeek R1 Distill Qwen 32B config/profile comparison, safetensors index, linked-object HEAD checks, and direct shard header range reads. Bound weights separate the exact 19.328804864 GB resident AWQ/BF16/F16 package from 17.771669504 GB ordinary text-decode swept traffic; BF16 full-context KV is retained from the served dtype and disabled sliding-window config." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/casperhansen/deepseek-r1-distill-qwen-32b-awq" }, { "label": "Model config", "url": "https://huggingface.co/casperhansen/deepseek-r1-distill-qwen-32b-awq/raw/e20a4933e66aa5eccc8270489f5aeab17f90b888/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/casperhansen/deepseek-r1-distill-qwen-32b-awq" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/casperhansen/deepseek-r1-distill-qwen-32b-awq/raw/e20a4933e66aa5eccc8270489f5aeab17f90b888/model.safetensors.index.json" } ] }, { "id": "hugging-quants/Meta-Llama-3.1-70B-Instruct-AWQ-INT4", "name": "Meta-Llama-3.1-70B-Instruct-AWQ-INT4 4-bit", "short_name": "Meta Llama 3.1 70B Instruct AWQ INT4", "license": "llama3.1", "hf_pipeline_tag": "text-generation", "hf_downloads": 282835, "tags": [ "transformers", "safetensors", "llama", "text-generation", "llama-3.1", "meta", "autoawq", "conversational", "en", "de", "fr", "it", "pt", "hi", "es", "th", "license:llama3.1", "text-generation-inference", "endpoints_compatible", "4-bit", "awq", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "llama (LlamaForCausalLM)", "total_params_b": 70.553706496, "active_params_b": 69.503033344, "layers": 80, "attention_heads": 64, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "AWQ 4-bit", "weight_bytes_per_param": 0.5636526760540158, "resident_weight_gb": 39.767785472, "active_weight_gb": 37.666439168, "kv_alloc_gb_per_1k_tokens": 0.32768, "kv_read_gb_per_1k_tokens": 0.32768, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, model card, pinned served AWQ config, gated Meta base API metadata, safetensors index, linked-object HEAD checks, and direct shard header range reads. Bound weights separate the exact 39.767785472 GB resident AWQ/F16 package from 37.666439168 GB ordinary text-decode swept traffic; FP16 full-context KV is retained from the served config dtype." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/hugging-quants/Meta-Llama-3.1-70B-Instruct-AWQ-INT4" }, { "label": "Model config", "url": "https://huggingface.co/hugging-quants/Meta-Llama-3.1-70B-Instruct-AWQ-INT4/raw/2123003760781134cfc31124aa6560a45b491fdf/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/meta-llama/Llama-3.1-70B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/hugging-quants/Meta-Llama-3.1-70B-Instruct-AWQ-INT4" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/hugging-quants/Meta-Llama-3.1-70B-Instruct-AWQ-INT4/raw/2123003760781134cfc31124aa6560a45b491fdf/model.safetensors.index.json" } ] }, { "id": "unsloth/gpt-oss-20b-GGUF", "name": "gpt-oss-20b-GGUF GGUF", "short_name": "gpt oss 20b", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 405984, "tags": [ "transformers", "gguf", "gpt_oss", "text-generation", "openai", "unsloth", "base_model:openai/gpt-oss-20b", "base_model:quantized:openai/gpt-oss-20b", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "MoE transformer", "detail": "gpt_oss (GptOssForCausalLM)", "total_params_b": 20.914757184, "active_params_b": 3.6, "layers": 24, "attention_heads": 64, "kv_heads": 8, "head_dim": 64, "max_context_tokens": 131072, "routed_experts": 32, "routed_experts_per_token": 4, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "GGUF F16/MXFP4", "weight_bytes_per_param": 0.6594692468412451, "resident_weight_gb": 13.792639168, "active_weight_gb": 3.714836736, "kv_alloc_gb_per_1k_tokens": 0.049152, "kv_read_gb_per_1k_tokens": 0.049152, "notes": "Manually corrected on 2026-07-05 from current Hugging Face API metadata, model card, served config, OpenAI base config comparison, selected linked-object HEAD checks, and a direct GGUF header/tensor-index range read. The authoritative profile targets the API-selected gpt-oss-20b-F16.gguf artifact: 13.792639168 GB resident file size, 12.621363456 GB main resident tensor traffic, 1.171275712 GB resident-only input embedding plus GGUF overhead, and 3.714836736 GB single-session ordinary text-decode swept traffic." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/gpt-oss-20b-GGUF" }, { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/gpt-oss-20b-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/unsloth/gpt-oss-20b-GGUF/raw/d449b42d93e1c2c7bda5312f5c25c8fb91dfa9b4/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/openai/gpt-oss-20b" }, { "label": "Selected F16 GGUF artifact", "url": "https://huggingface.co/unsloth/gpt-oss-20b-GGUF/resolve/d449b42d93e1c2c7bda5312f5c25c8fb91dfa9b4/gpt-oss-20b-F16.gguf" } ] }, { "id": "casperhansen/llama-3.3-70b-instruct-awq", "name": "llama-3.3-70b-instruct-awq 4-bit", "short_name": "llama 3.3 70b instruct", "license": "llama3.3", "hf_pipeline_tag": "text-generation", "hf_downloads": 318125, "tags": [ "transformers", "safetensors", "llama", "text-generation", "conversational", "en", "fr", "it", "pt", "hi", "es", "th", "de", "base_model:meta-llama/Llama-3.1-70B", "base_model:quantized:meta-llama/Llama-3.1-70B", "license:llama3.3", "text-generation-inference", "endpoints_compatible", "4-bit", "awq", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "llama (LlamaForCausalLM)", "total_params_b": 70.553706496, "active_params_b": 69.503033344, "layers": 80, "attention_heads": 64, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "AWQ INT4", "weight_bytes_per_param": 0.5636526760540158, "resident_weight_gb": 39.767785472, "active_weight_gb": 37.666439168, "kv_alloc_gb_per_1k_tokens": 0.32768, "kv_read_gb_per_1k_tokens": 0.32768, "notes": "Manually corrected on 2026-07-05 from current Hugging Face API metadata, model card, served AWQ config, gated Llama 3.3 base API metadata, safetensors index, linked-object HEAD checks, and direct range-read safetensors shard headers. Legacy fallback values mirror the audited profile: exact stored AWQ resident bytes, ordinary text-decode swept bytes excluding only model.embed_tokens.weight, and FP16 full-context KV. The README frontmatter/API cardData list base_model meta-llama/Llama-3.1-70B, but the README body and served config _name_or_path identify the executable artifact as Llama 3.3 70B Instruct." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/casperhansen/llama-3.3-70b-instruct-awq" }, { "label": "Hugging Face model card", "url": "https://huggingface.co/casperhansen/llama-3.3-70b-instruct-awq" }, { "label": "Model config", "url": "https://huggingface.co/casperhansen/llama-3.3-70b-instruct-awq/raw/64d255621f40b42adaf6d1f32a47e1d4534c0f14/config.json" }, { "label": "Base model API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.3-70B-Instruct" }, { "label": "Safetensors index", "url": "https://huggingface.co/casperhansen/llama-3.3-70b-instruct-awq/raw/64d255621f40b42adaf6d1f32a47e1d4534c0f14/model.safetensors.index.json" } ] }, { "id": "google/gemma-4-E4B-it-assistant", "name": "Gemma 4 E4B IT Assistant Drafter BF16", "short_name": "Gemma 4 E4B IT Assistant", "license": "apache-2.0", "hf_pipeline_tag": "any-to-any", "hf_downloads": 334689, "tags": [ "transformers", "safetensors", "gemma4_assistant", "text-generation", "any-to-any", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 0.078779908, "active_params_b": 0.078779908, "layers": 4, "attention_heads": 4, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "I64/BF16", "weight_bytes_per_param": 2.0199652936888426, "resident_weight_gb": 0.15913268, "active_weight_gb": 0.15913268, "kv_alloc_gb_per_1k_tokens": 0.008192, "kv_read_gb_per_1k_tokens": 0.008192, "notes": "Manually reviewed on 2026-07-05 from current Hugging Face API metadata, served config, generation config, model card, and direct safetensors header range read. This repo is a Gemma 4 MTP speculative drafter used as assistant_model with google/gemma-4-E4B-it. The profile records the exact 0.159132680 GB resident BF16/I64 drafter artifact and intentionally disables production throughput because Bounds Engine v1 lacks paired target+assistant speculative decoding." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-4-E4B-it-assistant" }, { "label": "Model config", "url": "https://huggingface.co/google/gemma-4-E4B-it-assistant/raw/1754649e3d145ddb62afd52e7204d5a755653aec/config.json" }, { "label": "Generation config", "url": "https://huggingface.co/google/gemma-4-E4B-it-assistant/raw/1754649e3d145ddb62afd52e7204d5a755653aec/generation_config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-E4B-it-assistant" }, { "label": "Target model card", "url": "https://huggingface.co/google/gemma-4-E4B-it" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/google/gemma-4-E4B-it-assistant/resolve/1754649e3d145ddb62afd52e7204d5a755653aec/model.safetensors" } ] }, { "id": "cyankiwi/Qwen3-30B-A3B-Thinking-2507-AWQ-4bit", "name": "Qwen3-30B-A3B-Thinking-2507-AWQ-4bit 4-bit", "short_name": "Qwen3 30B A3B Thinking 2507 AWQ 4bit", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 420409, "tags": [ "transformers", "safetensors", "qwen3_moe", "text-generation", "conversational", "arxiv:2402.17463", "arxiv:2407.02490", "arxiv:2501.15383", "arxiv:2404.06654", "arxiv:2505.09388", "base_model:Qwen/Qwen3-30B-A3B-Thinking-2507", "base_model:quantized:Qwen/Qwen3-30B-A3B-Thinking-2507", "license:apache-2.0", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_moe (Qwen3MoeForCausalLM)", "total_params_b": 30.532122624, "active_params_b": 3.3, "layers": 48, "attention_heads": 32, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "compressed-tensors AWQ INT4", "weight_bytes_per_param": 0.5924123756067358, "resident_weight_gb": 18.087607296, "active_weight_gb": 2.17676288, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "notes": "Manually corrected on 2026-07-05 from current Hugging Face API metadata, model card, pinned compressed-tensors config, BF16 base config comparison, recipe.yaml, safetensors index, and direct range-read safetensors shard headers. Legacy fallback values mirror the audited profile: exact 18.087607296 GB stored resident bytes, 1.157528576 GB fixed ordinary text traffic, 0.127404288 GB per routed expert group, and BF16 full-context KV. The authoritative profile supersedes the generated ideal 4-bit resident and active estimates." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3-30B-A3B-Thinking-2507-AWQ-4bit" }, { "label": "Hugging Face model card", "url": "https://huggingface.co/cyankiwi/Qwen3-30B-A3B-Thinking-2507-AWQ-4bit" }, { "label": "Model config", "url": "https://huggingface.co/cyankiwi/Qwen3-30B-A3B-Thinking-2507-AWQ-4bit/raw/4f1525a21bef5cbd32587a1f392ce75b281b633e/config.json" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Thinking-2507/raw/144afc2f379b542fdd4e85a1fcd5e1f79112d95d/config.json" }, { "label": "Safetensors index", "url": "https://huggingface.co/cyankiwi/Qwen3-30B-A3B-Thinking-2507-AWQ-4bit/raw/4f1525a21bef5cbd32587a1f392ce75b281b633e/model.safetensors.index.json" }, { "label": "Quantization recipe", "url": "https://huggingface.co/cyankiwi/Qwen3-30B-A3B-Thinking-2507-AWQ-4bit/raw/4f1525a21bef5cbd32587a1f392ce75b281b633e/recipe.yaml" } ] }, { "id": "Qwen/Qwen2.5-Coder-7B", "name": "Qwen2.5-Coder-7B BF16", "short_name": "Qwen2.5 Coder 7B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 450017, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "code", "qwen", "qwen-coder", "codeqwen", "conversational", "en", "arxiv:2409.12186", "arxiv:2309.00071", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-7B", "base_model:finetune:Qwen/Qwen2.5-7B", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 7.615616512, "active_params_b": 7.070619136, "layers": 28, "attention_heads": 28, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 15.231233024, "active_weight_gb": 14.141238272, "kv_alloc_gb_per_1k_tokens": 0.057344, "kv_read_gb_per_1k_tokens": 0.057344, "notes": "Manually corrected on 2026-07-05 from current Hugging Face API metadata, model card, pinned served config, Qwen2.5-7B base config comparison, safetensors index, and direct range-read safetensors shard headers. Legacy fallback values mirror the audited profile: exact 15.231233024 GB resident BF16 package, 14.141238272 GB ordinary text-decode swept traffic, 1.089994752 GB resident-only input embedding, and BF16 full-context KV. The authoritative profile supersedes this legacy adapter data for production bounds." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Coder-7B" }, { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B/raw/0396a76181e127dfc13e5c5ec48a8cee09938b02/config.json" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen2.5-7B/raw/d149729398750b98c0af14eb82c78cfe92750796/config.json" }, { "label": "Safetensors index", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B/raw/0396a76181e127dfc13e5c5ec48a8cee09938b02/model.safetensors.index.json" } ] }, { "id": "cyankiwi/Qwen3-30B-A3B-Instruct-2507-AWQ-4bit", "name": "Qwen3-30B-A3B-Instruct-2507-AWQ-4bit 4-bit", "short_name": "Qwen3 30B A3B Instruct 2507 AWQ 4bit", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 534049, "tags": [ "transformers", "safetensors", "qwen3_moe", "text-generation", "conversational", "arxiv:2402.17463", "arxiv:2407.02490", "arxiv:2501.15383", "arxiv:2404.06654", "arxiv:2505.09388", "base_model:Qwen/Qwen3-30B-A3B-Instruct-2507", "base_model:quantized:Qwen/Qwen3-30B-A3B-Instruct-2507", "license:apache-2.0", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_moe (Qwen3MoeForCausalLM)", "total_params_b": 30.532122624, "active_params_b": 3.3, "layers": 48, "attention_heads": 32, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "compressed-tensors AWQ INT4", "weight_bytes_per_param": 0.5924123756067358, "resident_weight_gb": 18.087607296, "active_weight_gb": 2.17676288, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "notes": "Manually corrected on 2026-07-05 from current Hugging Face API metadata, model card, pinned compressed-tensors config, BF16 base config comparison, recipe.yaml, safetensors index, and direct range-read safetensors shard headers. Legacy fallback values mirror the audited profile: exact 18.087607296 GB stored resident bytes, 1.157528576 GB fixed ordinary text traffic, 0.127404288 GB per routed expert group, and BF16 full-context KV. The authoritative profile supersedes the generated ideal 4-bit resident and active estimates." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3-30B-A3B-Instruct-2507-AWQ-4bit" }, { "label": "Hugging Face model card", "url": "https://huggingface.co/cyankiwi/Qwen3-30B-A3B-Instruct-2507-AWQ-4bit" }, { "label": "Model config", "url": "https://huggingface.co/cyankiwi/Qwen3-30B-A3B-Instruct-2507-AWQ-4bit/raw/876f886e383033237a04d8634b5af1cde0818290/config.json" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507/raw/0d7cf23991f47feeb3a57ecb4c9cee8ea4a17bfe/config.json" }, { "label": "Safetensors index", "url": "https://huggingface.co/cyankiwi/Qwen3-30B-A3B-Instruct-2507-AWQ-4bit/raw/876f886e383033237a04d8634b5af1cde0818290/model.safetensors.index.json" }, { "label": "Quantization recipe", "url": "https://huggingface.co/cyankiwi/Qwen3-30B-A3B-Instruct-2507-AWQ-4bit/raw/876f886e383033237a04d8634b5af1cde0818290/recipe.yaml" } ] }, { "id": "deepseek-ai/DeepSeek-R1-Distill-Qwen-7B", "name": "DeepSeek-R1-Distill-Qwen-7B BF16", "short_name": "DeepSeek R1 Distill Qwen 7B", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 329897, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "conversational", "arxiv:2501.12948", "license:mit", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 7.615616512, "active_params_b": 7.070619136, "layers": 28, "attention_heads": 28, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 15.231233024, "active_weight_gb": 14.141238272, "kv_alloc_gb_per_1k_tokens": 0.057344, "kv_read_gb_per_1k_tokens": 0.057344, "notes": "Manually corrected on 2026-07-05 from current Hugging Face API metadata, pinned DeepSeek model card, pinned served config, Qwen2.5 Math 7B base config comparison, safetensors index, and direct range-read safetensors shard headers. Legacy fallback values mirror the audited profile: exact 15.231233024 GB resident BF16 package, 14.141238272 GB ordinary text-decode swept traffic, 1.089994752 GB resident-only input embedding, and BF16 full-context KV. The authoritative profile supersedes this legacy adapter data for production bounds." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B" }, { "label": "Hugging Face model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B" }, { "label": "Model config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B/raw/916b56a44061fd5cd7d6a8fb632557ed4f724f60/config.json" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen2.5-Math-7B/raw/b101308fe89651ea5ce025f25317fea6fc07e96e/config.json" }, { "label": "Safetensors index", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B/raw/916b56a44061fd5cd7d6a8fb632557ed4f724f60/model.safetensors.index.json" } ] }, { "id": "nvidia/Phi-4-reasoning-plus-NVFP4", "name": "Phi-4-reasoning-plus-NVFP4 NVFP4", "short_name": "Phi 4 reasoning plus NVFP4", "license": "mit", "hf_pipeline_tag": "unknown", "hf_downloads": 472305, "tags": [ "Model Optimizer", "safetensors", "phi3", "nvidia", "ModelOpt", "Phi4", "quantized", "FP4", "fp4", "base_model:microsoft/Phi-4-reasoning-plus", "base_model:quantized:microsoft/Phi-4-reasoning-plus", "license:mit", "8-bit", "modelopt", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "phi3 (Phi3ForCausalLM)", "total_params_b": 14.6595072, "active_params_b": 14.14570496, "layers": 40, "attention_heads": 40, "kv_heads": 10, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "ModelOpt NVFP4 + FP8 scales + BF16 fixed tensors", "weight_bytes_per_param": 0.5, "resident_weight_gb": 9.723752, "active_weight_gb": 8.69614752, "kv_alloc_gb_per_1k_tokens": 0.1024, "kv_read_gb_per_1k_tokens": 0.1024, "notes": "Manually corrected on 2026-07-05 from current Hugging Face API metadata, NVIDIA model card, pinned served config, ModelOpt quantization config, Microsoft base config comparison, safetensors index, and direct range-read safetensors shard headers. Legacy fallback values mirror the audited profile: exact 9.723752 GB stored resident package, 8.69614752 GB ordinary text-decode swept traffic, 1.02760448 GB resident-only input embedding, and FP8 full-context KV. The authoritative profile supersedes the generated simple 0.5-byte NVFP4 resident estimate and BF16 KV estimate." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/Phi-4-reasoning-plus-NVFP4" }, { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/Phi-4-reasoning-plus-NVFP4" }, { "label": "Model config", "url": "https://huggingface.co/nvidia/Phi-4-reasoning-plus-NVFP4/raw/cb950fd61cdbfa8e0f467ace3087c7d32ea8a47b/config.json" }, { "label": "ModelOpt quantization config", "url": "https://huggingface.co/nvidia/Phi-4-reasoning-plus-NVFP4/raw/cb950fd61cdbfa8e0f467ace3087c7d32ea8a47b/hf_quant_config.json" }, { "label": "Base model config", "url": "https://huggingface.co/microsoft/Phi-4-reasoning-plus/raw/69baf8528e1bcf05f475034d9e5dd32875ed125f/config.json" }, { "label": "Safetensors index", "url": "https://huggingface.co/nvidia/Phi-4-reasoning-plus-NVFP4/raw/cb950fd61cdbfa8e0f467ace3087c7d32ea8a47b/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen3.5-122B-A10B-GPTQ-Int4", "name": "Qwen3.5-122B-A10B-GPTQ-Int4 4-bit", "short_name": "Qwen3.5 122B A10B GPTQ Int4", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 252591, "tags": [ "transformers", "safetensors", "qwen3_5_moe", "image-text-to-text", "conversational", "license:apache-2.0", "endpoints_compatible", "4-bit", "gptq", "deploy:azure", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_5_moe_text", "total_params_b": 125.086497008, "active_params_b": 9.008456448, "layers": 48, "attention_heads": 32, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "GPTQ Int4 packed I32 + F16 scales + BF16/F32 fixed tensors", "weight_bytes_per_param": 0.6303164604806022, "resident_weight_gb": 78.844078048, "active_weight_gb": 12.66288384, "kv_alloc_gb_per_1k_tokens": 0.024576, "kv_read_gb_per_1k_tokens": 0.024576, "kv_alloc_fixed_gb": 0.154533888, "kv_read_fixed_gb": 0.154533888, "expert_param_b": 0.3755510871785092, "fixed_param_b": 17.085315487063053, "shared_expert_param_b": 0, "notes": "Manually corrected on 2026-07-05 from current Hugging Face API metadata, pinned GPTQ config, BF16 base config comparison, safetensors index, direct range-read safetensors shard headers, linked-object HEAD checks, and representative tensor dtype inspection. Legacy fallback values mirror the audited profile: exact 78.844078048 GB stored resident package, 10.769155584 GB fixed text traffic, 0.236716032 GB per routed expert, 8 routed experts plus 1 shared expert, BF16 attention KV, and fixed BF16-convolution/F32-SSM DeltaNet recurrent state. The authoritative profile supersedes the generated 62.5433 GB resident estimate, missing shared-expert count, and all-layers full-context KV estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B-GPTQ-Int4" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B-GPTQ-Int4/raw/30cd92cba9707a9aba09d1e490ed4b66b78e9606/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-122B-A10B-GPTQ-Int4" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B/raw/dc4d348443bc740c68e2d77492492c11606384d5/config.json" }, { "label": "Safetensors index", "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B-GPTQ-Int4/resolve/30cd92cba9707a9aba09d1e490ed4b66b78e9606/model.safetensors.index.json" } ] }, { "id": "casperhansen/deepseek-r1-distill-llama-70b-awq", "name": "deepseek-r1-distill-llama-70b-awq 4-bit", "short_name": "deepseek r1 distill llama 70b", "license": "mit", "hf_pipeline_tag": "unknown", "hf_downloads": 423417, "tags": [ "safetensors", "llama", "license:mit", "4-bit", "awq", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "llama (LlamaForCausalLM)", "total_params_b": 70.553706496, "active_params_b": 69.503033344, "layers": 80, "attention_heads": 64, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "AWQ INT4 packed I32 + F16 scales + BF16 fixed tensors", "weight_bytes_per_param": 0.5636526760540158, "resident_weight_gb": 39.767785472, "active_weight_gb": 37.666439168, "kv_alloc_gb_per_1k_tokens": 0.32768, "kv_read_gb_per_1k_tokens": 0.32768, "notes": "Manually corrected on 2026-07-05 from current Hugging Face API metadata, served AWQ config, DeepSeek source snapshot API/config comparison, safetensors index, linked-object HEAD checks, and direct range-read safetensors shard headers. Legacy fallback values mirror the audited profile: exact stored AWQ resident bytes, ordinary text-decode swept bytes excluding only model.embed_tokens.weight, and BF16 full-context KV. The package config and generation config set use_cache false by default, but the architecture and runtime implementation support standard cached decode." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/casperhansen/deepseek-r1-distill-llama-70b-awq" }, { "label": "Hugging Face model card", "url": "https://huggingface.co/casperhansen/deepseek-r1-distill-llama-70b-awq" }, { "label": "Model config", "url": "https://huggingface.co/casperhansen/deepseek-r1-distill-llama-70b-awq/raw/a1ab7653aae77fbabc536cbcbac5bb2e2fb5354f/config.json" }, { "label": "DeepSeek source snapshot config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B/raw/07a264a567ba0863a4ab34fdb3c2b8a54e0bb494/config.json" }, { "label": "Safetensors index", "url": "https://huggingface.co/casperhansen/deepseek-r1-distill-llama-70b-awq/resolve/a1ab7653aae77fbabc536cbcbac5bb2e2fb5354f/model.safetensors.index.json" } ] }, { "id": "RedHatAI/gemma-4-31B-it-NVFP4", "name": "gemma-4-31B-it-NVFP4 NVFP4", "short_name": "gemma 4 31B it NVFP4", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 342927, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "fp4", "vllm", "llm-compressor", "compressed-tensors", "conversational", "base_model:google/gemma-4-31B-it", "base_model:quantized:google/gemma-4-31B-it", "license:apache-2.0", "endpoints_compatible", "8-bit", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 32.682372656, "active_params_b": 32.682372656, "layers": 60, "attention_heads": 32, "kv_heads": 16, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "compressed-tensors NVFP4 + BF16/F8/U8 side tensors", "weight_bytes_per_param": 0.5, "resident_weight_gb": 23.26508228, "active_weight_gb": 19.29502292, "kv_alloc_gb_per_1k_tokens": 0.83968, "kv_read_gb_per_1k_tokens": 0.83968, "notes": "Manual correction on 2026-07-05 from live Hugging Face API metadata, served compressed-tensors config, recipe.yaml, Google base config comparison, linked safetensors object metadata, and a direct safetensors header range read. The authoritative bounds profile uses exact stored bytes: 23.26508228 GB resident, 19.29502292 GB swept ordinary text traffic, 3.97005936 GB resident-only embedding and vision traffic, separate lm_head output projection despite tie_word_embeddings true, and layered Gemma 4 BF16 KV because kv_cache_scheme is null. Legacy per-1k KV fields show the short-context unsaturated coefficient; long-context calculation must use the profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/RedHatAI/gemma-4-31B-it-NVFP4" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/RedHatAI/gemma-4-31B-it-NVFP4" }, { "label": "Pinned model config", "url": "https://huggingface.co/RedHatAI/gemma-4-31B-it-NVFP4/raw/34484c941404ed94fb46f04d7e0b04ac592b49a8/config.json" }, { "label": "Pinned base config", "url": "https://huggingface.co/google/gemma-4-31B-it/raw/3548789868c5356dbf307c98e6f609007b82b3eb/config.json" }, { "label": "Safetensors linked-object and header audit", "url": "https://huggingface.co/RedHatAI/gemma-4-31B-it-NVFP4/resolve/34484c941404ed94fb46f04d7e0b04ac592b49a8/model.safetensors" } ] }, { "id": "cyankiwi/gemma-4-12B-it-AWQ-INT4", "name": "gemma-4-12B-it-AWQ-INT4 4-bit", "short_name": "gemma 4 12B it AWQ INT4", "license": "apache-2.0", "hf_pipeline_tag": "any-to-any", "hf_downloads": 364002, "tags": [ "transformers", "safetensors", "gemma4_unified", "image-text-to-text", "any-to-any", "base_model:google/gemma-4-12B-it", "base_model:quantized:google/gemma-4-12B-it", "license:apache-2.0", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_unified_text", "total_params_b": 12.6409776, "active_params_b": 12.588597696, "layers": 48, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "compressed-tensors AWQ INT4 + F32 side tensors", "weight_bytes_per_param": 0.8877410671149358, "resident_weight_gb": 11.221914944, "active_weight_gb": 11.012395328, "kv_alloc_gb_per_1k_tokens": 0.671744, "kv_read_gb_per_1k_tokens": 0.671744, "notes": "Manual correction on 2026-07-05 from live Hugging Face API metadata, HF CLI repo info/file listing, served compressed-tensors config, Google base config comparison, linked safetensors object metadata, and a direct safetensors header range read. The authoritative bounds profile uses exact stored bytes: 11.221914944 GB resident, 11.012395328 GB swept ordinary text traffic, 0.209519616 GB resident-only multimodal projection traffic, tied embedding/logit projection charged once because no lm_head tensor is stored, 131072 served max context, and layered Gemma 4 FP32 KV from text_config dtype float32. Legacy per-1k KV fields show the short-context unsaturated coefficient; long-context calculation must use the profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/cyankiwi/gemma-4-12B-it-AWQ-INT4" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/cyankiwi/gemma-4-12B-it-AWQ-INT4" }, { "label": "Pinned model config", "url": "https://huggingface.co/cyankiwi/gemma-4-12B-it-AWQ-INT4/raw/aa32360daab742921c34e01a503c024fc8c6a510/config.json" }, { "label": "Pinned base config", "url": "https://huggingface.co/google/gemma-4-12B-it/raw/5926caa4ec0cac5cbfadaf4077420520de1d5205/config.json" }, { "label": "Safetensors linked-object and header audit", "url": "https://huggingface.co/cyankiwi/gemma-4-12B-it-AWQ-INT4/resolve/aa32360daab742921c34e01a503c024fc8c6a510/model.safetensors" } ] }, { "id": "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8", "name": "NVIDIA-Nemotron-3-Nano-30B-A3B-FP8 FP8", "short_name": "NVIDIA Nemotron 3 Nano 30B A3B FP8", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 332474, "tags": [ "transformers", "safetensors", "nemotron_h", "text-generation", "nvidia", "pytorch", "conversational", "custom_code", "en", "es", "fr", "de", "ja", "it", "dataset:nvidia/Nemotron-Pretraining-Code-v1", "dataset:nvidia/Nemotron-CC-v2", "dataset:nvidia/Nemotron-Pretraining-SFT-v1", "dataset:nvidia/Nemotron-CC-Math-v1", "dataset:nvidia/Nemotron-Pretraining-Code-v2", "dataset:nvidia/Nemotron-Pretraining-Specialized-v1", "dataset:nvidia/Nemotron-CC-v2.1", "dataset:nvidia/Nemotron-CC-Code-v1", "dataset:nvidia/Nemotron-Pretraining-Dataset-sample", "dataset:nvidia/Nemotron-Competitive-Programming-v1", "dataset:nvidia/Nemotron-Math-v2", "dataset:nvidia/Nemotron-Agentic-v1", "dataset:nvidia/Nemotron-Math-Proofs-v1", "dataset:nvidia/Nemotron-Instruction-Following-Chat-v1", "dataset:nvidia/Nemotron-Science-v1", "dataset:nvidia/Nemotron-3-Nano-RL-Training-Blend", "arxiv:2512.20848", "arxiv:2512.20856", "base_model:nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16", "base_model:quantized:nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16", "license:other", "eval-results", "endpoints_compatible", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "nemotron_h (NemotronHForCausalLM)", "total_params_b": 31.577952224, "active_params_b": 3.227755528, "layers": 52, "attention_heads": 32, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 6, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "ModelOpt FP8 + BF16 exclusions", "weight_bytes_per_param": 1, "resident_weight_gb": 32.679949312, "active_weight_gb": 3.977397408, "kv_alloc_gb_per_1k_tokens": 0.003072, "kv_read_gb_per_1k_tokens": 0.003072, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, pinned served config, BF16 config comparison, ModelOpt quantization config, model card serving notes, custom runtime review, safetensors index, and direct range-read safetensors shard headers. Legacy fallback values mirror the audited profile: exact stored resident bytes, fixed always-on hybrid/shared-expert traffic, expected distinct routed experts, FP8 attention KV, and fixed BF16-convolution/F32-SSM Mamba recurrent state. The authoritative profile supersedes the generated 31.5779 GB resident estimate and all-layer KV estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8" }, { "label": "Pinned model config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8/raw/f8dc1c0afee92f44417695b4f5ddca9afc95ea58/config.json" }, { "label": "Pinned quantization config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8/raw/f8dc1c0afee92f44417695b4f5ddca9afc95ea58/hf_quant_config.json" }, { "label": "Pinned BF16 config comparison", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/raw/cbd3fa9f933d55ef16a84236559f4ee2a0526848/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8/resolve/f8dc1c0afee92f44417695b4f5ddca9afc95ea58/model.safetensors.index.json" } ] }, { "id": "cyankiwi/Qwen3-VL-2B-Instruct-AWQ-4bit", "name": "Qwen3-VL-2B-Instruct-AWQ-4bit 4-bit", "short_name": "Qwen3 VL 2B Instruct AWQ 4bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 351144, "tags": [ "transformers", "safetensors", "qwen3_vl", "image-text-to-text", "conversational", "arxiv:2505.09388", "arxiv:2502.13923", "arxiv:2409.12191", "arxiv:2308.12966", "base_model:Qwen/Qwen3-VL-2B-Instruct", "base_model:quantized:Qwen/Qwen3-VL-2B-Instruct", "license:apache-2.0", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_vl_text", "total_params_b": 2.171572616, "active_params_b": 1.76461556, "layers": 28, "attention_heads": 16, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "compressed-tensors AWQ INT4 + BF16 side tensors", "weight_bytes_per_param": 1.0265456248505207, "resident_weight_gb": 2.229218368, "active_weight_gb": 1.415304256, "kv_alloc_gb_per_1k_tokens": 0.114688, "kv_read_gb_per_1k_tokens": 0.114688, "notes": "Manual correction on 2026-07-05 from live Hugging Face API metadata, HF CLI repo info/file listing, served compressed-tensors config, Qwen base config comparison, recipe.yaml, linked safetensors object metadata, and a direct safetensors header range read. The authoritative bounds profile uses exact stored bytes: 2.229218368 GB resident, 1.415304256 GB swept ordinary text traffic, 0.813914112 GB resident-only visual traffic, tied embedding/logit projection charged once because no lm_head tensor is stored, 262144 served max context, and full-context Qwen3-VL BF16 KV. Legacy per-1k KV fields mirror the full-context text coefficient." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/cyankiwi/Qwen3-VL-2B-Instruct-AWQ-4bit" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3-VL-2B-Instruct-AWQ-4bit" }, { "label": "Pinned model config", "url": "https://huggingface.co/cyankiwi/Qwen3-VL-2B-Instruct-AWQ-4bit/raw/66408a5298047012e0059208c6306a3a64e597af/config.json" }, { "label": "Pinned base config", "url": "https://huggingface.co/Qwen/Qwen3-VL-2B-Instruct/raw/89644892e4d85e24eaac8bacfd4f463576704203/config.json" }, { "label": "Pinned AWQ recipe", "url": "https://huggingface.co/cyankiwi/Qwen3-VL-2B-Instruct-AWQ-4bit/raw/66408a5298047012e0059208c6306a3a64e597af/recipe.yaml" }, { "label": "Safetensors linked-object and header audit", "url": "https://huggingface.co/cyankiwi/Qwen3-VL-2B-Instruct-AWQ-4bit/resolve/66408a5298047012e0059208c6306a3a64e597af/model.safetensors" } ] }, { "id": "DavidAU/Qwen3.6-27B-Heretic-Uncensored-FINETUNE-NEO-CODE-Di-IMatrix-MAX-GGUF", "name": "Qwen3.6-27B-Heretic-Uncensored-FINETUNE-NEO-CODE-Di-IMatrix-MAX-GGUF GGUF", "short_name": "Qwen3.6 27B Heretic Uncensored FINETUNE NEO CODE", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 235458, "tags": [ "gguf", "unsloth", "heretic", "uncensored", "abliterated", "fine tune", "creative", "creative writing", "fiction writing", "plot generation", "sub-plot generation", "story generation", "scene continue", "storytelling", "fiction story", "science fiction", "romance", "all genres", "story", "writing", "vivid prosing", "vivid writing", "fiction", "roleplaying", "bfloat16", "all use cases", "image-text-to-text", "en", "zh", "base_model:DavidAU/Qwen3.6-27B-Heretic2-Uncensored-Finetune-Thinking", "base_model:quantized:DavidAU/Qwen3.6-27B-Heretic2-Uncensored-Finetune-Thinking", "license:apache-2.0", "endpoints_compatible", "region:us", "imatrix", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 26.895998464, "active_params_b": 25.624600064, "layers": 64, "attention_heads": 24, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q8_0 HIGH", "weight_bytes_per_param": 1.1104381121963467, "resident_weight_gb": 29.86634176, "active_weight_gb": 28.504487936, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, model card, full-precision base config, Qwen3.6 parent config comparison, HF CLI linked-object sizes, and a direct GGUF header range read. The authoritative bounds profile targets the API-selected Qwen3.6-27B-NEO-CODE-HERE-2T-OT-HIGH-Q8_0.gguf artifact: 29.866341760 GB resident file size, 28.504487936 GB swept ordinary text traffic, 1.361853824 GB resident-only token embedding plus GGUF overhead, 262144 context, and Qwen3.6 hybrid FP16 KV plus DeltaNet state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/DavidAU/Qwen3.6-27B-Heretic-Uncensored-FINETUNE-NEO-CODE-Di-IMatrix-MAX-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/DavidAU/Qwen3.6-27B-Heretic-Uncensored-FINETUNE-NEO-CODE-Di-IMatrix-MAX-GGUF" }, { "label": "Pinned base config", "url": "https://huggingface.co/DavidAU/Qwen3.6-27B-Heretic2-Uncensored-Finetune-Thinking/raw/2bca4e1caef0db358b478988446b95993226965d/config.json" }, { "label": "Pinned Qwen3.6 parent config", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json" }, { "label": "Selected Q8_0 GGUF header audit", "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" } ] }, { "id": "nvidia/Llama-3.1-8B-Instruct-FP8", "name": "Llama-3.1-8B-Instruct-FP8 FP8", "short_name": "Llama 3.1 8B Instruct FP8", "license": "llama3.1", "hf_pipeline_tag": "text-generation", "hf_downloads": 240873, "tags": [ "transformers", "safetensors", "llama", "text-generation", "conversational", "base_model:meta-llama/Llama-3.1-8B-Instruct", "base_model:quantized:meta-llama/Llama-3.1-8B-Instruct", "license:llama3.1", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "llama (LlamaForCausalLM)", "total_params_b": 8.03026176, "active_params_b": 7.504925184, "layers": 32, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "FP8", "weight_bytes_per_param": 1, "resident_weight_gb": 9.081202688, "active_weight_gb": 8.030529536, "kv_alloc_gb_per_1k_tokens": 0.065536, "kv_read_gb_per_1k_tokens": 0.065536, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, NVIDIA model card, pinned public config, pinned ModelOpt quantization config, safetensors index, linked-object HEAD checks, and direct shard header range reads. The profile records the exact 9.081202688 GB resident ModelOpt FP8 package, the 8.030529536 GB ordinary text-decode swept subset, the 1.050673152 GB resident-only input embedding, and FP8 KV cache traffic from hf_quant_config.json." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/Llama-3.1-8B-Instruct-FP8" }, { "label": "Model config", "url": "https://huggingface.co/nvidia/Llama-3.1-8B-Instruct-FP8/raw/42d9515ebd69eea3a87351d079c671c3c5ff0a31/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/Llama-3.1-8B-Instruct-FP8" }, { "label": "ModelOpt quantization config", "url": "https://huggingface.co/nvidia/Llama-3.1-8B-Instruct-FP8/raw/42d9515ebd69eea3a87351d079c671c3c5ff0a31/hf_quant_config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/nvidia/Llama-3.1-8B-Instruct-FP8/raw/42d9515ebd69eea3a87351d079c671c3c5ff0a31/model.safetensors.index.json" } ] }, { "id": "apolo13x/Qwen3.5-9B-NVFP4", "name": "Qwen3.5-9B-NVFP4 NVFP4", "short_name": "Qwen3.5 9B NVFP4", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 334541, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "quantized", "nvfp4", "fp4", "4-bit", "vllm", "llm-compressor", "conversational", "dataset:nvidia/Nemotron-Post-Training-Dataset-v2", "base_model:Qwen/Qwen3.5-9B", "base_model:quantized:Qwen/Qwen3.5-9B", "license:apache-2.0", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 9.653104368, "active_params_b": 7.936684544, "layers": 32, "attention_heads": 16, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "NVFP4", "weight_bytes_per_param": 1.280462906935391, "resident_weight_gb": 12.36044208, "active_weight_gb": 8.927602432, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "notes": "Manual correction on 2026-07-05 from current Hugging Face API metadata, model card, pinned compressed-tensors config, pinned quantization recipe, base Qwen3.5 config comparison, safetensors index, linked-object HEAD checks, and direct shard header range reads. The profile records the exact 12.360442080 GB resident llm-compressor NVFP4 package, the 8.927602432 GB ordinary text-decode swept subset, the 3.432839648 GB resident-only input embedding plus visual plus MTP subset, and BF16 hybrid Qwen3.5 KV/state traffic because kv_cache_scheme is null." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/apolo13x/Qwen3.5-9B-NVFP4" }, { "label": "Model config", "url": "https://huggingface.co/apolo13x/Qwen3.5-9B-NVFP4/raw/83fa19092744f3819c129ad41d21190cfa67bae1/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-9B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/apolo13x/Qwen3.5-9B-NVFP4" }, { "label": "Pinned Qwen3.5 9B base config", "url": "https://huggingface.co/Qwen/Qwen3.5-9B/raw/c202236235762e1c871ad0ccb60c8ee5ba337b9a/config.json" }, { "label": "Quantization recipe", "url": "https://huggingface.co/apolo13x/Qwen3.5-9B-NVFP4/raw/83fa19092744f3819c129ad41d21190cfa67bae1/recipe.yaml" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/apolo13x/Qwen3.5-9B-NVFP4/raw/83fa19092744f3819c129ad41d21190cfa67bae1/model.safetensors.index.json" } ] }, { "id": "RedHatAI/gemma-3-27b-it-quantized.w4a16", "name": "gemma-3-27b-it-quantized.w4a16 I64/I32/BF16", "short_name": "gemma 3 27b it quantized.w4a16", "license": "gemma", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 323684, "tags": [ "transformers", "safetensors", "gemma3", "image-text-to-text", "vllm", "vision", "w4a16", "conversational", "base_model:google/gemma-3-27b-it", "base_model:quantized:google/gemma-3-27b-it", "license:gemma", "text-generation-inference", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense multimodal transformer", "detail": "gemma3_text (Gemma3ForConditionalGeneration)", "total_params_b": 29.042024916, "active_params_b": 27.209334372, "layers": 62, "attention_heads": 32, "kv_heads": 16, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "compressed-tensors GPTQ W4A16", "weight_bytes_per_param": 0.6778629485010252, "resident_weight_gb": 19.68651264, "active_weight_gb": 16.021131552, "kv_alloc_gb_per_1k_tokens": 0.08192, "kv_read_gb_per_1k_tokens": 0.08192, "notes": "Manually audited on 2026-07-05 from current HF API metadata, served compressed-tensors config, model card metadata, quantization recipe, safetensors index, direct safetensors header range reads, and Gemma 3 local/global attention documentation. downloads=323684. Exact tensor payload bytes: 19.686512640 GB resident, 16.021131552 GB swept ordinary text traffic, and 3.665381088 GB resident-only input embedding/vision/projector bytes. KV fields show only the full-context component per 1K tokens; the audited profile also charges 52 sliding-window layers capped at 1024 tokens." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/RedHatAI/gemma-3-27b-it-quantized.w4a16" }, { "label": "Model config", "url": "https://huggingface.co/RedHatAI/gemma-3-27b-it-quantized.w4a16/raw/2b537554d6c6f6368945e8df4e5fb7bbbb5d56c9/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-3-27b-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/RedHatAI/gemma-3-27b-it-quantized.w4a16" }, { "label": "Quantization recipe", "url": "https://huggingface.co/RedHatAI/gemma-3-27b-it-quantized.w4a16/raw/2b537554d6c6f6368945e8df4e5fb7bbbb5d56c9/recipe.yaml" }, { "label": "Safetensors index", "url": "https://huggingface.co/RedHatAI/gemma-3-27b-it-quantized.w4a16/raw/2b537554d6c6f6368945e8df4e5fb7bbbb5d56c9/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen3-VL-30B-A3B-Instruct-FP8", "name": "Qwen3-VL-30B-A3B-Instruct-FP8 FP8", "short_name": "Qwen3 VL 30B A3B Instruct FP8", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 515645, "tags": [ "transformers", "safetensors", "qwen3_vl_moe", "image-text-to-text", "conversational", "arxiv:2505.09388", "arxiv:2502.13923", "arxiv:2409.12191", "arxiv:2308.12966", "base_model:Qwen/Qwen3-VL-30B-A3B-Instruct", "base_model:quantized:Qwen/Qwen3-VL-30B-A3B-Instruct", "license:apache-2.0", "endpoints_compatible", "fp8", "deploy:azure", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_vl_moe_text", "total_params_b": 31.0725788, "active_params_b": 3.042033664, "layers": 48, "attention_heads": 32, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "FP8 E4M3 + BF16/F32 side tensors", "weight_bytes_per_param": 1.0379508064518932, "resident_weight_gb": 32.251808224, "active_weight_gb": 3.366490112, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "notes": "Manually audited on 2026-07-05 from current HF API metadata, model card, pinned FP8 config, BF16 base config comparison, official Transformers Qwen3 VL MoE implementation, safetensors index, and direct range-read shard headers. downloads=515645. Exact tensor payload bytes: 32.251808224 GB resident, 30.552215552 GB main text/logit resident, 1.699592672 GB resident-only visual/input embedding bytes, 1.554108416 GB fixed non-expert traffic, and 0.226547712 GB per routed expert. active_weight_gb is single-sequence ordinary text-decode traffic: fixed plus 8 routed experts. The audited profile supersedes these legacy adapter fields for expected-distinct expert batching." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct-FP8" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct-FP8/raw/d9748a51ae66354c4dad665aab2c71f26cf2c8cd/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-VL-30B-A3B-Instruct-FP8" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct/raw/9c4b90e1e4ba969fd3b5378b57d966d725f1b86c/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct-FP8/resolve/d9748a51ae66354c4dad665aab2c71f26cf2c8cd/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen3-32B-AWQ", "name": "Qwen3-32B-AWQ 4-bit", "short_name": "Qwen3 32B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 1602159, "tags": [ "transformers", "safetensors", "qwen3", "text-generation", "conversational", "arxiv:2309.00071", "arxiv:2505.09388", "base_model:Qwen/Qwen3-32B", "base_model:quantized:Qwen/Qwen3-32B", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "4-bit", "awq", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 32.762123264, "active_params_b": 31.984210944, "layers": 64, "attention_heads": 64, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "AWQ 4-bit + BF16 side tensors", "weight_bytes_per_param": 0.5898671014779806, "resident_weight_gb": 19.325298688, "active_weight_gb": 17.769474048, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Manually audited on 2026-07-05 from current HF API metadata, model card, pinned AWQ config, BF16 base config comparison, existing BF16 profile, safetensors index, linked-object HEAD checks, and direct range-read shard headers. downloads=1602159. Exact tensor payload bytes: 19.325298688 GB resident, 17.769474048 GB swept ordinary text traffic, and 1.555824640 GB resident-only input embedding bytes. The safetensors index metadata reports 19.338405888 GB, but direct headers and linked object sizes show 19.325298688 GB tensor payload; bounds use direct tensor spans." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-32B-AWQ" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-32B-AWQ/raw/0499c3ac83fdef8810b907a23894ba91e95eddd8/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-32B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-32B-AWQ" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3-32B/raw/9216db5781bf21249d130ec9da846c4624c16137/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Qwen/Qwen3-32B-AWQ/raw/0499c3ac83fdef8810b907a23894ba91e95eddd8/model.safetensors.index.json" } ] }, { "id": "nvidia/Llama-4-Scout-17B-16E-Instruct-FP8", "name": "Llama-4-Scout-17B-16E-Instruct-FP8 FP8", "short_name": "Llama 4 Scout 17B 16E Instruct FP8", "license": "other", "hf_pipeline_tag": "unknown", "hf_downloads": 334755, "tags": [ "Model Optimizer", "safetensors", "llama4", "nvidia", "modelopt", "quantized", "FP8", "base_model:meta-llama/Llama-4-Scout-17B-16E-Instruct", "base_model:quantized:meta-llama/Llama-4-Scout-17B-16E-Instruct", "license:other", "region:us" ], "architecture": { "type": "Unknown gated multimodal MoE", "detail": "gated ModelOpt FP8 config unavailable", "total_params_b": 108.641793536, "active_params_b": 108.641793536, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "FP8 + BF16 side tensors", "weight_bytes_per_param": 1.0271104944067775, "resident_weight_gb": 111.587126272, "active_weight_gb": 111.587126272, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manual fail-closed correction on 2026-07-05 from current Hugging Face API metadata, rendered model card, and direct raw-file/HF CLI access checks. The API records 108.641793536B safetensors parameters at commit 3c0175a5a4a66f8e139faf93ddeb7d2741ff3560: 2.945332736B BF16 and 105.696460800B F8_E4M3, or 111.587126272 GB if charged by stored dtype. Raw config, README, generation_config.json, hf_quant_config.json, model.safetensors.index.json, and hf download config.json all require accepted-gate access in this environment. Production bounds use the unsupported profile and must not infer KV geometry, MoE routing, context, or swept traffic from this row." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/Llama-4-Scout-17B-16E-Instruct-FP8" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/Llama-4-Scout-17B-16E-Instruct-FP8" }, { "label": "Gated config access check", "url": "https://huggingface.co/nvidia/Llama-4-Scout-17B-16E-Instruct-FP8/raw/3c0175a5a4a66f8e139faf93ddeb7d2741ff3560/config.json" }, { "label": "Gated tensor index access check", "url": "https://huggingface.co/nvidia/Llama-4-Scout-17B-16E-Instruct-FP8/raw/3c0175a5a4a66f8e139faf93ddeb7d2741ff3560/model.safetensors.index.json" }, { "label": "Base model card", "url": "https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct" } ] }, { "id": "lmstudio-community/DeepSeek-R1-0528-Qwen3-8B-MLX-4bit", "name": "DeepSeek-R1-0528-Qwen3-8B-MLX-4bit 4-bit", "short_name": "DeepSeek R1 0528 Qwen3 8B MLX 4bit", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 318926, "tags": [ "mlx", "safetensors", "qwen3", "text-generation", "conversational", "base_model:deepseek-ai/DeepSeek-R1-0528-Qwen3-8B", "base_model:quantized:deepseek-ai/DeepSeek-R1-0528-Qwen3-8B", "license:mit", "4-bit", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 8.19073536, "active_params_b": 7.568405504, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "MLX 4-bit affine", "weight_bytes_per_param": 0.5625540942883057, "resident_weight_gb": 4.607731712, "active_weight_gb": 4.257671168, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, pinned served MLX config, pinned BF16 DeepSeek source config comparison, model card, safetensors index, and direct single-file safetensors header range read. Bound weights separate the exact 4.607731712 GB resident MLX tensor payload from the 4.257671168 GB ordinary text-decode swept traffic; the audited profile supersedes the generated ideal 4-bit estimate and retains BF16 full-context KV from the served Qwen3 config." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/DeepSeek-R1-0528-Qwen3-8B-MLX-4bit" }, { "label": "Model config", "url": "https://huggingface.co/lmstudio-community/DeepSeek-R1-0528-Qwen3-8B-MLX-4bit/raw/1b6065ad0cd7256f662c9ef6fd0ed2dfd06e286f/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-0528-Qwen3-8B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/DeepSeek-R1-0528-Qwen3-8B-MLX-4bit" }, { "label": "Pinned BF16 source config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-0528-Qwen3-8B/raw/6e8885a6ff5c1dc5201574c8fd700323f23c25fa/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/lmstudio-community/DeepSeek-R1-0528-Qwen3-8B-MLX-4bit/raw/1b6065ad0cd7256f662c9ef6fd0ed2dfd06e286f/model.safetensors.index.json" } ] }, { "id": "nvidia/Cosmos-Reason2-8B", "name": "Cosmos-Reason2-8B BF16", "short_name": "Cosmos Reason2 8B", "license": "other", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 183046, "tags": [ "cosmos", "safetensors", "qwen3_vl", "nvidia", "conversational", "image-text-to-text", "arxiv:2503.06800", "arxiv:2406.10721", "arxiv:2603.18178", "arxiv:2312.14115", "base_model:Qwen/Qwen3-VL-8B-Instruct", "base_model:finetune:Qwen/Qwen3-VL-8B-Instruct", "license:other", "region:us" ], "architecture": { "type": "Unknown gated Qwen3-VL finetune", "detail": "qwen3_vl_text (gated config unavailable)", "total_params_b": 8.767123696, "active_params_b": 8.767123696, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 17.534247392, "active_weight_gb": 17.534247392, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manual fail-closed correction on 2026-07-05 from current Hugging Face API metadata and direct raw-file/HF CLI access checks. The API records 8.767123696B BF16 safetensors parameters at commit a9fae2cf89dc64db96b12860417f0eb403013bb9, but raw config, README, generation_config.json, preprocessor configs, model.safetensors.index.json, and hf download config.json all require accepted-gate access in this environment. Production bounds use the unsupported profile and must not infer KV geometry, Qwen3-VL context, text/vision split, or swept traffic from this row." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/Cosmos-Reason2-8B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/Cosmos-Reason2-8B" }, { "label": "Gated config access check", "url": "https://huggingface.co/nvidia/Cosmos-Reason2-8B/raw/a9fae2cf89dc64db96b12860417f0eb403013bb9/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct" } ] }, { "id": "Qwen/Qwen3-Omni-30B-A3B-Thinking", "name": "Qwen3-Omni-30B-A3B-Thinking BF16", "short_name": "Qwen3 Omni 30B A3B Thinking", "license": "other", "hf_pipeline_tag": "any-to-any", "hf_downloads": 347831, "tags": [ "transformers", "safetensors", "qwen3_omni_moe", "text-to-audio", "multimodal", "any-to-any", "en", "license:other", "endpoints_compatible", "region:us" ], "architecture": { "type": "Multimodal MoE transformer", "detail": "qwen3_omni_moe Thinking checkpoint; ordinary text decode uses the 48-layer Thinker text MoE", "total_params_b": 31.719205488, "active_params_b": 3.04212992, "layers": 48, "attention_heads": 32, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 65536, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 63.438410976, "active_weight_gb": 6.08425984, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "expert_param_b": 0.226492416, "fixed_param_b": 1.230190592, "notes": "Manually audited on 2026-07-05 from current Hugging Face API metadata, model card, pinned served config, safetensors index, and direct shard-header range reads. The Thinking checkpoint contains the Thinker package without Talker/code2wav tensors. Bound weights separate the exact 63.438410976 GB resident package from ordinary text-decode traffic through fixed Thinker text weights plus expected distinct routed experts; the authoritative profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Thinking" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Thinking/raw/2f443cfc4c54b14a815c0e2bb9a9d6cbcd9a748b/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-Omni-30B-A3B-Thinking" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Thinking/resolve/2f443cfc4c54b14a815c0e2bb9a9d6cbcd9a748b/model.safetensors.index.json" } ] }, { "id": "janhq/Jan-v3.5-4B-gguf", "name": "Jan-v3.5-4B-gguf GGUF", "short_name": "Jan v3.5 4B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 304639, "tags": [ "transformers", "gguf", "math", "identity", "text-generation", "en", "base_model:janhq/Jan-v3-4B-base-instruct", "base_model:quantized:janhq/Jan-v3-4B-base-instruct", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 4.411424256, "active_params_b": 4.022468096, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q3_K_L", "weight_bytes_per_param": 0.5456091358083152, "resident_weight_gb": 2.406913376, "active_weight_gb": 2.233828864, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Audited profile-backed row from live Hugging Face metadata on 2026-07-06; downloads=304639; selected artifact Jan-v3.5-4B-Q3_K_L.gguf; data quality: API metadata, model card, source config/index, base config comparison, GGUF HEAD checks, and direct GGUF header tensor index. Resident weight uses linked file size; active weight uses blk.* plus output.weight plus output_norm.weight tensor spans." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/janhq/Jan-v3.5-4B-gguf" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/janhq/Jan-v3.5-4B-gguf/resolve/2b0fcab0066198e78f7c004a0815179fbce13985/Jan-v3.5-4B-Q3_K_L.gguf" }, { "label": "Source model config", "url": "https://huggingface.co/janhq/Jan-v3.5-4B/raw/53509d2d88feb0a1fccadf26e185383fc6a75d8e/config.json" }, { "label": "Source model safetensors index", "url": "https://huggingface.co/janhq/Jan-v3.5-4B/raw/53509d2d88feb0a1fccadf26e185383fc6a75d8e/model.safetensors.index.json" }, { "label": "Base model config", "url": "https://huggingface.co/janhq/Jan-v3-4B-base-instruct/raw/f331713ecaef85c9da95fe5b63131ddb0b8c6744/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/janhq/Jan-v3-4B-base-instruct" } ] }, { "id": "google/gemma-4-E2B-it-qat-q4_0-gguf", "name": "gemma-4-E2B-it-qat-q4_0-gguf GGUF 4-bit", "short_name": "gemma 4 E2B it qat q4 0", "license": "apache-2.0", "hf_pipeline_tag": "any-to-any", "hf_downloads": 346498, "tags": [ "transformers", "gguf", "any-to-any", "base_model:google/gemma-4-E2B-it-qat-q4_0-unquantized", "base_model:quantized:google/gemma-4-E2B-it-qat-q4_0-unquantized", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 4.628569635, "active_params_b": 2.279759395, "layers": 35, "attention_heads": 8, "kv_heads": 1, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q4_0 QAT", "weight_bytes_per_param": 0.7236607367148318, "resident_weight_gb": 3.349514112, "active_weight_gb": 1.406942304, "kv_alloc_gb_per_1k_tokens": 0.03584, "kv_read_gb_per_1k_tokens": 0.03584, "notes": "Audited profile-backed row from live Hugging Face metadata on 2026-07-06; downloads=346498; selected artifact gemma-4-E2B_q4_0-it.gguf; data quality: API metadata, model card, QAT source config, Gemma 4 E2B config comparison, GGUF HEAD checks, Transformers PLE documentation, and direct GGUF header tensor index. Resident weight uses linked file size; active weight excludes the resident-only per-layer token embedding lookup table." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-4-E2B-it-qat-q4_0-gguf" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-E2B-it-qat-q4_0-gguf" }, { "label": "Selected GGUF tensor index", "url": "https://huggingface.co/google/gemma-4-E2B-it-qat-q4_0-gguf/resolve/1894d1fc0a19d86697abd40483f5983c867df03f/gemma-4-E2B_q4_0-it.gguf" }, { "label": "Base model config", "url": "https://huggingface.co/google/gemma-4-E2B-it-qat-q4_0-unquantized/raw/d40c88a04b9e3aae9cc9d6f63389e786bd20fba4/config.json" }, { "label": "Gemma 4 PLE documentation", "url": "https://huggingface.co/docs/transformers/model_doc/gemma4" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-E2B-it-qat-q4_0-unquantized" } ] }, { "id": "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8", "name": "NVIDIA-Nemotron-3-Super-120B-A12B-FP8 FP8", "short_name": "NVIDIA Nemotron 3 Super 120B A12B FP8", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 265465, "tags": [ "transformers", "safetensors", "nemotron_h", "text-generation", "nvidia", "pytorch", "nemotron-3", "latent-moe", "mtp", "conversational", "custom_code", "en", "fr", "es", "it", "de", "ja", "zh", "dataset:nvidia/nemotron-post-training-v3", "dataset:nvidia/nemotron-pre-training-datasets", "arxiv:2512.20848", "arxiv:2512.20856", "license:other", "eval-results", "endpoints_compatible", "modelopt", "deploy:azure", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "nemotron_h (NemotronHForCausalLM)", "total_params_b": 123.611012096, "active_params_b": 12.233387008, "layers": 88, "attention_heads": 32, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 512, "routed_experts_per_token": 22, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "ModelOpt FP8 + BF16 exclusions", "weight_bytes_per_param": 1, "resident_weight_gb": 128.334890312, "active_weight_gb": 13.47771296, "kv_alloc_gb_per_1k_tokens": 0.004096, "kv_read_gb_per_1k_tokens": 0.004096, "kv_alloc_fixed_gb": 0.17104896, "kv_read_fixed_gb": 0.17104896, "expert_param_b": 0.2202016, "fixed_param_b": 8.63327776, "shared_expert_param_b": 0, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, pinned served config, BF16 config comparison, ModelOpt quantization config, model card serving notes, custom runtime review, safetensors index, and direct range-read safetensors shard headers. Legacy fallback values mirror the audited profile: exact stored resident bytes, fixed always-on hybrid/shared-expert traffic, expected distinct routed experts, FP8 attention KV, and fixed BF16-convolution/F32-SSM Mamba recurrent state. The authoritative profile supersedes the generated 123.611 GB resident estimate and all-layer KV estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8" }, { "label": "Pinned served config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8/raw/7d7e5797b8a3c7abbab54033b6004e93e8b6bc91/config.json" }, { "label": "Pinned ModelOpt quantization config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8/raw/7d7e5797b8a3c7abbab54033b6004e93e8b6bc91/hf_quant_config.json" }, { "label": "Pinned BF16 config comparison", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16/raw/d51eab0d1f979ebc26b546e634a04f450d99158e/config.json" }, { "label": "Custom runtime code", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8/raw/7d7e5797b8a3c7abbab54033b6004e93e8b6bc91/modeling_nemotron_h.py" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8/resolve/7d7e5797b8a3c7abbab54033b6004e93e8b6bc91/model.safetensors.index.json" } ] }, { "id": "moonshotai/Kimi-VL-A3B-Instruct", "name": "Kimi-VL-A3B-Instruct BF16", "short_name": "Kimi VL A3B", "license": "mit", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 332831, "tags": [ "transformers", "safetensors", "kimi_vl", "feature-extraction", "agent", "video", "screenspot", "long-context", "image-text-to-text", "conversational", "custom_code", "arxiv:2504.07491", "base_model:moonshotai/Moonlight-16B-A3B", "base_model:finetune:moonshotai/Moonlight-16B-A3B", "license:mit", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "kimi_vl with DeepseekV3 language model", "total_params_b": 16.407657776, "active_params_b": 2.5792336, "layers": 27, "attention_heads": 16, "kv_heads": 16, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": 64, "routed_experts_per_token": 6, "shared_experts_per_token": 2 }, "adapter": { "kind": "moe", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 32.815315552, "active_weight_gb": 5.1584672, "kv_alloc_gb_per_1k_tokens": 0.27648, "kv_read_gb_per_1k_tokens": 0.27648, "expert_param_b": 0.224919552, "fixed_param_b": 1.229716288, "shared_expert_param_b": 0, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, pinned served config, Moonlight base config comparison, model card, custom runtime review, safetensors index, and direct range-read safetensors shard headers. Legacy fallback values mirror the audited profile: exact BF16 resident bytes, resident-only vision/projector/input-embedding tensors, fixed always-on language/shared-expert traffic, expected distinct routed experts, and expanded BF16 K/V cache from the custom DeepSeekV3 attention path." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/moonshotai/Kimi-VL-A3B-Instruct" }, { "label": "Pinned served config", "url": "https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct/raw/398eede0903cd983a2bfa0cc634e9ac1d843f375/config.json" }, { "label": "Pinned Moonlight base config comparison", "url": "https://huggingface.co/moonshotai/Moonlight-16B-A3B/raw/476b36a473d4467f94469414bef6cee75c9c8172/config.json" }, { "label": "Custom runtime code", "url": "https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct/raw/398eede0903cd983a2bfa0cc634e9ac1d843f375/modeling_kimi_vl.py" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct/resolve/398eede0903cd983a2bfa0cc634e9ac1d843f375/model.safetensors.index.json" } ] }, { "id": "google/gemma-4-26B-A4B-it-qat-q4_0-gguf", "name": "gemma-4-26B-A4B-it-qat-q4_0-gguf GGUF 4-bit", "short_name": "gemma 4 26B A4B it qat q4 0", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 338211, "tags": [ "transformers", "gguf", "image-text-to-text", "base_model:google/gemma-4-26B-A4B-it-qat-q4_0-unquantized", "base_model:quantized:google/gemma-4-26B-A4B-it-qat-q4_0-unquantized", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "MoE transformer", "detail": "gemma4_text", "total_params_b": 25.233142046, "active_params_b": 3.8, "layers": 30, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "GGUF Q4_0 QAT", "weight_bytes_per_param": 0.5722379485550018, "resident_weight_gb": 14.43936144, "active_weight_gb": 2.380070008, "kv_alloc_gb_per_1k_tokens": 0.20992, "kv_read_gb_per_1k_tokens": 0.20992, "expert_param_b": 0.1753855022258341, "fixed_param_b": 2.7561473194544823, "shared_expert_param_b": 0, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, pinned model card, public QAT config, non-QAT config comparison, linked GGUF file sizes, and a direct GGUF tensor-index range read. The authoritative bounds profile uses the selected gemma-4-26B_q4_0-it.gguf artifact: 14.439361440 GB resident, 1.577172088 GB fixed traffic, 0.100362240 GB per routed expert, 8 routed experts per token, 128 routed experts, 1 shared expert charged in fixed traffic, and layered Gemma 4 hybrid KV. Legacy per-1k KV fields show the short-context unsaturated coefficient; long-context calculation must use the profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it-qat-q4_0-gguf" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-26B-A4B-it-qat-q4_0-gguf" }, { "label": "Pinned QAT base config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it-qat-q4_0-unquantized/raw/641f184470aa8554ae7957599a624badc2bf4e57/config.json" }, { "label": "Pinned non-QAT config comparison", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it/raw/20da991ab4afab98e8f910c4a2e8f4fbefc404ad/config.json" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it-qat-q4_0-gguf/resolve/dfc00409adc70be497fee9c90bfe76b3ee130f2e/gemma-4-26B_q4_0-it.gguf" } ] }, { "id": "RedHatAI/gemma-4-31B-it-FP8-Dynamic", "name": "gemma-4-31B-it-FP8-Dynamic FP8", "short_name": "gemma 4 31B it FP8 Dynamic", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 329630, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "fp8", "vllm", "llm-compressor", "compressed-tensors", "conversational", "base_model:google/gemma-4-31B-it", "base_model:quantized:google/gemma-4-31B-it", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 32.682372656, "active_params_b": 32.682372656, "layers": 60, "attention_heads": 32, "kv_heads": 16, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "FP8/BF16", "weight_bytes_per_param": 1.017915946989623, "resident_weight_gb": 33.267908312, "active_weight_gb": 32.11642124, "kv_alloc_gb_per_1k_tokens": 0.83968, "kv_read_gb_per_1k_tokens": 0.83968, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, pinned model card, pinned compressed-tensors config, base config comparison, linked object metadata, and direct safetensors shard-header range reads. Direct headers found 33.267908312 GB resident package bytes, 32.116421240 GB ordinary text-decode swept language bytes including the tied embedding/output projection, and 1.151487072 GB resident-only vision bytes. The model.safetensors.index total_size is stale/unquantized for this artifact, so shard headers are authoritative. Legacy per-1k KV fields show the first-1K unsaturated Gemma 4 hybrid BF16 KV coefficient; long-context calculation must use the profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/RedHatAI/gemma-4-31B-it-FP8-dynamic" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/RedHatAI/gemma-4-31B-it-FP8-dynamic" }, { "label": "Pinned compressed-tensors config", "url": "https://huggingface.co/RedHatAI/gemma-4-31B-it-FP8-dynamic/raw/c991bf46174b1523aa4208d602dfb072287719a7/config.json" }, { "label": "Pinned base config comparison", "url": "https://huggingface.co/google/gemma-4-31B-it/raw/3548789868c5356dbf307c98e6f609007b82b3eb/config.json" }, { "label": "Safetensors index and shard-header audit", "url": "https://huggingface.co/RedHatAI/gemma-4-31B-it-FP8-dynamic/raw/c991bf46174b1523aa4208d602dfb072287719a7/model.safetensors.index.json" } ] }, { "id": "allenai/OLMo-2-0425-1B", "name": "OLMo-2-0425-1B F32", "short_name": "OLMo 2 0425 1B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 202403, "tags": [ "transformers", "safetensors", "olmo2", "text-generation", "en", "arxiv:2501.00656", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "olmo2 (Olmo2ForCausalLM)", "total_params_b": 1.484916736, "active_params_b": 1.484916736, "layers": 16, "attention_heads": 16, "kv_heads": 16, "head_dim": 128, "max_context_tokens": 4096, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "F32", "weight_bytes_per_param": 4, "resident_weight_gb": 5.939666944, "active_weight_gb": 5.11758336, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, pinned model card, pinned config, linked object metadata, safetensors index, and direct safetensors shard-header range reads. Direct headers found 5.939666944 GB resident F32 tensor payload, 5.117583360 GB ordinary text-decode swept traffic, and 0.822083584 GB resident-only input embedding. The audited profile charges FP32 KV because the served config records torch_dtype float32 and all safetensors payloads are F32." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/allenai/OLMo-2-0425-1B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/allenai/OLMo-2-0425-1B" }, { "label": "Pinned model config", "url": "https://huggingface.co/allenai/OLMo-2-0425-1B/raw/a1847dff35000b4271fa70afc5db10fd29fedbdf/config.json" }, { "label": "Safetensors index and shard-header audit", "url": "https://huggingface.co/allenai/OLMo-2-0425-1B/raw/a1847dff35000b4271fa70afc5db10fd29fedbdf/model.safetensors.index.json" } ] }, { "id": "google/gemma-2-9b-it", "name": "gemma-2-9b-it BF16", "short_name": "gemma 2 9b it", "license": "gemma", "hf_pipeline_tag": "text-generation", "hf_downloads": 247024, "tags": [ "transformers", "safetensors", "gemma2", "text-generation", "conversational", "arxiv:2009.03300", "arxiv:1905.07830", "arxiv:1911.11641", "arxiv:1904.09728", "arxiv:1905.10044", "arxiv:1907.10641", "arxiv:1811.00937", "arxiv:1809.02789", "arxiv:1911.01547", "arxiv:1705.03551", "arxiv:2107.03374", "arxiv:2108.07732", "arxiv:2110.14168", "arxiv:2009.11462", "arxiv:2101.11718", "arxiv:2110.08193", "arxiv:1804.09301", "arxiv:2109.07958", "arxiv:1804.06876", "arxiv:2103.03874", "arxiv:2304.06364", "arxiv:2206.04615", "arxiv:2203.09509", "base_model:google/gemma-2-9b", "base_model:finetune:google/gemma-2-9b", "license:gemma", "eval-results", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Gated decoder config unavailable", "detail": "gemma2 gated config unavailable", "total_params_b": 9.241705984, "active_params_b": 9.241705984, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 18.483411968, "active_weight_gb": 18.483411968, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manual fail-closed correction on 2026-07-06 from current Hugging Face API metadata and gated-access checks. The API records 9241705984 BF16 safetensors parameters, but raw config/model-card requests return 401 and hf download with the configured CLI token returns access denied because the repo requires approval. KV geometry, context length, layer count, head layout, and tied-embedding layout are intentionally not inferred. Production bounds use the unsupported profile until direct config and tensor evidence are available." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-2-9b-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-2-9b-it" }, { "label": "Gated config access check", "url": "https://huggingface.co/google/gemma-2-9b-it/raw/11c9b309abf73637e4b6f9a3fa1e92e615547819/config.json" } ] }, { "id": "sgl-project/DeepSeek-V4-Flash-FP8", "name": "DeepSeek-V4-Flash-FP8 FP8", "short_name": "DeepSeek V4 Flash FP8", "license": "mit", "hf_pipeline_tag": "unknown", "hf_downloads": 369947, "tags": [ "safetensors", "deepseek_v4", "deepseek-v4", "fp8", "quantized", "base_model:deepseek-ai/DeepSeek-V4-Flash", "base_model:quantized:deepseek-ai/DeepSeek-V4-Flash", "license:mit", "region:us" ], "architecture": { "type": "FP8 MoE transformer", "detail": "deepseek_v4 CSA/HCA MoE (DeepseekV4ForCausalLM)", "total_params_b": 284, "active_params_b": 13, "layers": 43, "attention_heads": 64, "kv_heads": 1, "head_dim": 512, "max_context_tokens": 1048576, "routed_experts": 256, "routed_experts_per_token": 6, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "FP8", "weight_bytes_per_param": 1, "resident_weight_gb": 294.03138004, "active_weight_gb": 15.724826204, "kv_alloc_gb_per_1k_tokens": 0.00688, "kv_read_gb_per_1k_tokens": 0.001504, "kv_alloc_fixed_gb": 0.017842176, "kv_read_fixed_gb": 0.016646144, "expert_param_b": 1.082394624, "shared_expert_param_b": 0, "fixed_param_b": 9.23045846, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, SGL model card, pinned served config, official DeepSeek V4 Flash profile/code review, linked-object metadata, safetensors index, and direct range-read shard headers. The SGL README states architecture/tokenizer/chat template/reference encoding are unchanged from the base model. Bound weights use SGL-specific all-FP8 stored bytes: 294.031380040 GB resident tensor payload, 9.230458460 GB fixed ordinary traffic, 1.082394624 GB per routed expert index, and 7.707897836 GB resident-only embed/MTP bytes. The index metadata.total_size is stale for this all-FP8 repack; shard headers and linked object sizes are authoritative." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/sgl-project/DeepSeek-V4-Flash-FP8" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/sgl-project/DeepSeek-V4-Flash-FP8" }, { "label": "Pinned model config", "url": "https://huggingface.co/sgl-project/DeepSeek-V4-Flash-FP8/raw/ae01d80c06cdfe30581edfd0e1c5449dc7ed7f17/config.json" }, { "label": "Safetensors index and shard-header audit", "url": "https://huggingface.co/sgl-project/DeepSeek-V4-Flash-FP8/raw/ae01d80c06cdfe30581edfd0e1c5449dc7ed7f17/model.safetensors.index.json" }, { "label": "Official DeepSeek V4 Flash inference review", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash/raw/60d8d70770c6776ff598c94bb586a859a38244f1/inference/model.py" } ] }, { "id": "lmstudio-community/Qwen3.6-27B-GGUF", "name": "Qwen3.6-27B-GGUF GGUF", "short_name": "Qwen3.6 27B", "license": "apache-2.0", "hf_pipeline_tag": "unknown", "hf_downloads": 331042, "tags": [ "gguf", "base_model:Qwen/Qwen3.6-27B", "base_model:quantized:Qwen/Qwen3.6-27B", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense multimodal transformer", "detail": "qwen3_5_text (API-selected Q4_K_M GGUF)", "total_params_b": 26.895998464, "active_params_b": 25.624600064, "layers": 64, "attention_heads": 24, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q4_K_M", "weight_bytes_per_param": 0.6152364563133252, "resident_weight_gb": 16.547398784, "active_weight_gb": 15.821244416, "kv_alloc_gb_per_1k_tokens": 0.065536, "kv_read_gb_per_1k_tokens": 0.065536, "kv_alloc_fixed_gb": 0.154927104, "kv_read_fixed_gb": 0.154927104, "notes": "Manually audited on 2026-07-06 from HF API commit 58c6607d9c4cae8b071b3781c73be633fb3dee36, model card, pinned Qwen base config, linked GGUF file sizes, and a GGUF tensor-index range read. The authoritative bounds profile uses the API-selected Qwen3.6-27B-Q4_K_M.gguf artifact: 16.547398784 GB resident, 15.821244416 GB swept ordinary text traffic, 0.726154368 GB resident-only token embedding plus GGUF metadata/header/file overhead, and Qwen3.6 hybrid KV plus fixed DeltaNet state. The separate mmproj sidecar is not included unless explicitly loaded." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-27B-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.6-27B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3.6-27B-GGUF" }, { "label": "Pinned base config", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-27B-GGUF/resolve/58c6607d9c4cae8b071b3781c73be633fb3dee36/Qwen3.6-27B-Q4_K_M.gguf" } ] }, { "id": "google/paligemma-3b-pt-224", "name": "paligemma-3b-pt-224 F32", "short_name": "paligemma 3b pt 224", "license": "gemma", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 278336, "tags": [ "transformers", "safetensors", "paligemma", "image-text-to-text", "arxiv:2310.09199", "arxiv:2303.15343", "arxiv:2403.08295", "arxiv:1706.03762", "arxiv:2010.11929", "arxiv:2209.06794", "arxiv:2209.04372", "arxiv:2103.01913", "arxiv:2205.12522", "arxiv:2110.11624", "arxiv:2108.03353", "arxiv:2010.04295", "arxiv:2401.06209", "arxiv:2305.10355", "arxiv:2203.10244", "arxiv:1810.12440", "arxiv:1905.13648", "arxiv:1608.00272", "arxiv:1908.04913", "arxiv:2407.07726", "license:gemma", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Unknown gated multimodal model", "detail": "PaliGemmaForConditionalGeneration (gated config unavailable)", "total_params_b": 2.92346648, "active_params_b": 2.92346648, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "F32", "weight_bytes_per_param": 4, "resident_weight_gb": 11.69386592, "active_weight_gb": 11.69386592, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Fail-closed on 2026-07-06 from current Hugging Face API metadata and gated-access checks. The API records 2923466480 F32 safetensors parameters at commit 35e4f46485b4d07967e7e9935bc3786aad50687c, but raw config, model.safetensors.index.json, ranged safetensors shard bytes, and hf download config.json all require manual approval in this environment. KV geometry, exact context, text/vision tensor split, tied embeddings, and swept decode traffic are intentionally unknown until direct gated evidence is available." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/paligemma-3b-pt-224" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/paligemma-3b-pt-224" }, { "label": "Pinned gated config access check", "url": "https://huggingface.co/google/paligemma-3b-pt-224/raw/35e4f46485b4d07967e7e9935bc3786aad50687c/config.json" } ] }, { "id": "Weni/Llama-Guard-3-8B-AWQ", "name": "Llama-Guard-3-8B-AWQ 4-bit", "short_name": "Llama Guard 3 8B", "license": "unknown", "hf_pipeline_tag": "unknown", "hf_downloads": 245818, "tags": [ "safetensors", "llama", "4-bit", "awq", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "llama (LlamaForCausalLM, AWQ GEMM)", "total_params_b": 8.030261248, "active_params_b": 7.504924672, "layers": 32, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "AWQ INT4", "weight_bytes_per_param": 0.7132837170679306, "resident_weight_gb": 5.727854592, "active_weight_gb": 4.67718144, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "notes": "Manually audited on 2026-07-06 from HF API commit 2157d713ed408803437e1a22f11d2c6dcb1034e2, served AWQ config, quant_config, gated base API metadata, linked-object HEAD checks, safetensors index, and direct shard-header range reads. The authoritative bounds profile uses exact stored AWQ/F16 resident bytes: 5.727854592 GB resident, 4.677181440 GB swept ordinary text traffic excluding only model.embed_tokens.weight, 1.050673152 GB resident-only input embedding, and FP16 full-context KV." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Weni/Llama-Guard-3-8B-AWQ" }, { "label": "Model config", "url": "https://huggingface.co/Weni/Llama-Guard-3-8B-AWQ/raw/main/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Weni/Llama-Guard-3-8B-AWQ" }, { "label": "Pinned model config", "url": "https://huggingface.co/Weni/Llama-Guard-3-8B-AWQ/raw/2157d713ed408803437e1a22f11d2c6dcb1034e2/config.json" }, { "label": "Safetensors index and shard-header audit", "url": "https://huggingface.co/Weni/Llama-Guard-3-8B-AWQ/raw/2157d713ed408803437e1a22f11d2c6dcb1034e2/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen3-30B-A3B-Instruct-2507-FP8", "name": "Qwen3-30B-A3B-Instruct-2507-FP8 FP8", "short_name": "Qwen3 30B A3B Instruct 2507 FP8", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 304783, "tags": [ "transformers", "safetensors", "qwen3_moe", "text-generation", "conversational", "arxiv:2505.09388", "base_model:Qwen/Qwen3-30B-A3B-Instruct-2507", "base_model:quantized:Qwen/Qwen3-30B-A3B-Instruct-2507", "license:apache-2.0", "endpoints_compatible", "fp8", "deploy:azure", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_moe (Qwen3MoeForCausalLM)", "total_params_b": 30.533947392, "active_params_b": 3.042033664, "layers": 48, "attention_heads": 32, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "FP8/BF16", "weight_bytes_per_param": 1, "resident_weight_gb": 31.170895872, "active_weight_gb": 3.366158336, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, model card, pinned served FP8 config, direct BF16 config comparison, safetensors index, linked-object HEAD checks, and direct shard-header range reads. Active traffic is fixed decode traffic plus 8 routed expert groups; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507-FP8" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507-FP8/raw/5a5a776300a41aaa681dd7ff0106608ef2bc90db/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-30B-A3B-Instruct-2507-FP8" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507-FP8/raw/5a5a776300a41aaa681dd7ff0106608ef2bc90db/model.safetensors.index.json" }, { "label": "BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507/raw/0d7cf23991f47feeb3a57ecb4c9cee8ea4a17bfe/config.json" } ] }, { "id": "unsloth/gemma-4-31B-it-qat-GGUF", "name": "gemma-4-31B-it-qat-GGUF GGUF 4-bit", "short_name": "gemma 4 31B it qat", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 407993, "tags": [ "transformers", "gguf", "gemma4", "image-text-to-text", "unsloth", "gemma", "google", "base_model:google/gemma-4-31B-it-qat-q4_0-unquantized", "base_model:quantized:google/gemma-4-31B-it-qat-q4_0-unquantized", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 30.697345596, "active_params_b": 30.697345596, "layers": 60, "attention_heads": 32, "kv_heads": 16, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q4_0 QAT", "weight_bytes_per_param": 0.5631649163259463, "resident_weight_gb": 17.287668064, "active_weight_gb": 17.271834864, "kv_alloc_gb_per_1k_tokens": 0.86016, "kv_read_gb_per_1k_tokens": 0.86016, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, pinned model card, packaged config, base config comparison, linked GGUF file sizes, and a direct GGUF tensor-index range read. The authoritative bounds profile uses the selected gemma-4-31B-it-qat-UD-Q4_K_XL.gguf artifact: 17.287668064 GB resident, 17.271834864 GB swept ordinary text traffic, tied token_embd.weight output projection, and layered Gemma 4 hybrid KV. Legacy per-1k KV fields show the short-context unsaturated coefficient; long-context calculation must use the profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/gemma-4-31B-it-qat-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/gemma-4-31B-it-qat-GGUF" }, { "label": "Pinned model config", "url": "https://huggingface.co/unsloth/gemma-4-31B-it-qat-GGUF/raw/365d657136993b4d7c40d868dd45ecb7a48e7ebf/config.json" }, { "label": "Pinned base config", "url": "https://huggingface.co/google/gemma-4-31B-it-qat-q4_0-unquantized/raw/4f926903562062220b3e54c1385c5ef2cd40bfd1/config.json" }, { "label": "GGUF linked-object and header audit", "url": "https://huggingface.co/unsloth/gemma-4-31B-it-qat-GGUF/resolve/365d657136993b4d7c40d868dd45ecb7a48e7ebf/gemma-4-31B-it-qat-UD-Q4_K_XL.gguf" } ] }, { "id": "lmstudio-community/DeepSeek-R1-0528-Qwen3-8B-MLX-8bit", "name": "DeepSeek-R1-0528-Qwen3-8B-MLX-8bit 8-bit", "short_name": "DeepSeek R1 0528 Qwen3 8B MLX 8bit", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 291320, "tags": [ "mlx", "safetensors", "qwen3", "text-generation", "conversational", "base_model:deepseek-ai/DeepSeek-R1-0528-Qwen3-8B", "base_model:quantized:deepseek-ai/DeepSeek-R1-0528-Qwen3-8B", "license:mit", "8-bit", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 8.19073536, "active_params_b": 7.568405504, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "MLX 8-bit affine", "weight_bytes_per_param": 1.0625352788836777, "resident_weight_gb": 8.70294528, "active_weight_gb": 8.041719808, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, pinned served MLX config, pinned BF16 DeepSeek source config comparison, model card, safetensors index, and direct two-shard safetensors header range reads. Bound weights separate the exact 8.702945280 GB resident MLX tensor payload from the 8.041719808 GB ordinary text-decode swept traffic; the audited profile supersedes the generated ideal 8-bit estimate and retains BF16 full-context KV from the served Qwen3 config." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/DeepSeek-R1-0528-Qwen3-8B-MLX-8bit" }, { "label": "Model config", "url": "https://huggingface.co/lmstudio-community/DeepSeek-R1-0528-Qwen3-8B-MLX-8bit/raw/e5242815680c2ed87df14b5a1deb76bd672dd69f/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-0528-Qwen3-8B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/DeepSeek-R1-0528-Qwen3-8B-MLX-8bit" }, { "label": "Pinned BF16 source config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-0528-Qwen3-8B/raw/6e8885a6ff5c1dc5201574c8fd700323f23c25fa/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/lmstudio-community/DeepSeek-R1-0528-Qwen3-8B-MLX-8bit/raw/e5242815680c2ed87df14b5a1deb76bd672dd69f/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen3.5-27B-GPTQ-Int4", "name": "Qwen3.5-27B-GPTQ-Int4 4-bit", "short_name": "Qwen3.5 27B GPTQ Int4", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 288886, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "conversational", "license:apache-2.0", "endpoints_compatible", "4-bit", "gptq", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 27.781427952, "active_params_b": 25.624600064, "layers": 64, "attention_heads": 24, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GPTQ Int4", "weight_bytes_per_param": 1.0883185652025984, "resident_weight_gb": 30.235043808, "active_weight_gb": 25.921388032, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Manually audited on 2026-07-06 from HF API commit 8f0c09f227ae570e79617c6d9172b59df9c16081, model card, served GPTQ config, BF16 base config comparison, safetensors index, linked-object HEAD checks, and direct safetensors header range reads. The authoritative bounds profile uses exact tensor payload bytes: 30.235043808 GB resident, 25.921388032 GB swept ordinary text traffic, and 4.313655776 GB resident-only input embedding/vision/MTP." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3.5-27B-GPTQ-Int4" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3.5-27B-GPTQ-Int4/raw/8f0c09f227ae570e79617c6d9172b59df9c16081/config.json" }, { "label": "Pinned BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3.5-27B/raw/fc05daec18b0a78c049392ed2e771dde82bdf654/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Qwen/Qwen3.5-27B-GPTQ-Int4/resolve/8f0c09f227ae570e79617c6d9172b59df9c16081/model.safetensors.index.json" } ] }, { "id": "lmstudio-community/Qwen3.6-35B-A3B-GGUF", "name": "Qwen3.6-35B-A3B-GGUF GGUF", "short_name": "Qwen3.6 35B A3B", "license": "apache-2.0", "hf_pipeline_tag": "unknown", "hf_downloads": 308995, "tags": [ "gguf", "base_model:Qwen/Qwen3.6-35B-A3B", "base_model:quantized:Qwen/Qwen3.6-35B-A3B", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "MoE transformer", "detail": "qwen35moe / qwen3_5_moe_text", "total_params_b": 34.660610688, "active_params_b": 3, "layers": 40, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "GGUF Q4_K_M", "weight_bytes_per_param": 0.6106862316574311, "resident_weight_gb": 21.166757728, "active_weight_gb": 1.975675392, "kv_alloc_gb_per_1k_tokens": 0.02048, "kv_read_gb_per_1k_tokens": 0.02048, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, model card, Qwen base config, selected linked-object HEAD checks, and direct GGUF header range read. The selected API artifact is Qwen3.6-35B-A3B-Q4_K_M.gguf at 21.166757728 GB resident; ordinary single-session swept traffic is 1.975675392 GB before KV/state reads. The audited profile adds the fixed DeltaNet state charge that these legacy per-1K KV fields cannot fully express." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-35B-A3B-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/995ad96eacd98c81ed38be0c5b274b04031597b0/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3.6-35B-A3B-GGUF" }, { "label": "Selected GGUF header audit", "url": "https://huggingface.co/lmstudio-community/Qwen3.6-35B-A3B-GGUF/resolve/68a34855558af61cbef0324d31f411be8a506b08/Qwen3.6-35B-A3B-Q4_K_M.gguf" } ] }, { "id": "google/gemma-4-E2B-it-qat-w4a16-ct", "name": "gemma-4-E2B-it-qat-w4a16-ct 4-bit", "short_name": "gemma 4 E2B it qat w4a16 ct", "license": "apache-2.0", "hf_pipeline_tag": "any-to-any", "hf_downloads": 381034, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "any-to-any", "base_model:google/gemma-4-E2B-it-qat-q4_0-unquantized", "base_model:quantized:google/gemma-4-E2B-it-qat-q4_0-unquantized", "license:apache-2.0", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 5.565602827, "active_params_b": 2.338410315, "layers": 35, "attention_heads": 8, "kv_heads": 1, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "W4A16 compressed-tensors", "weight_bytes_per_param": 1.4941739956105566, "resident_weight_gb": 8.315979014, "active_weight_gb": 1.86159399, "kv_alloc_gb_per_1k_tokens": 0.012435456, "kv_read_gb_per_1k_tokens": 0.029016064, "notes": "Audited profile-backed row from live Hugging Face metadata on 2026-07-06; downloads=381034; data quality: API metadata, model card, served compressed-tensors config, QAT source config comparison, upstream Transformers Gemma 4 implementation review, and direct safetensors header grouping. Resident weight uses exact stored tensor bytes; active weight excludes resident-only input embedding, PLE lookup, audio, and vision tensors. KV allocation/read summaries reflect Gemma 4 E2B shared-KV behavior at a 1k-token reference point; production bounds use the audited profile formulas." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-4-E2B-it-qat-w4a16-ct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-E2B-it-qat-w4a16-ct" }, { "label": "Pinned model config", "url": "https://huggingface.co/google/gemma-4-E2B-it-qat-w4a16-ct/raw/a50d96741aa53e1e0f1e8b6ea73230dce02b3341/config.json" }, { "label": "Pinned safetensors artifact", "url": "https://huggingface.co/google/gemma-4-E2B-it-qat-w4a16-ct/resolve/a50d96741aa53e1e0f1e8b6ea73230dce02b3341/model.safetensors" }, { "label": "QAT source config", "url": "https://huggingface.co/google/gemma-4-E2B-it-qat-q4_0-unquantized/raw/d40c88a04b9e3aae9cc9d6f63389e786bd20fba4/config.json" } ] }, { "id": "Qwen/Qwen2.5-VL-3B-Instruct-AWQ", "name": "Qwen2.5-VL-3B-Instruct-AWQ 4-bit", "short_name": "Qwen2.5 VL 3B Instruct", "license": "qwen-research", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 176189, "tags": [ "transformers", "safetensors", "qwen2_5_vl", "image-text-to-text", "multimodal", "conversational", "en", "arxiv:2309.00071", "arxiv:2409.12191", "arxiv:2308.12966", "base_model:Qwen/Qwen2.5-VL-3B-Instruct", "base_model:quantized:Qwen/Qwen2.5-VL-3B-Instruct", "text-generation-inference", "endpoints_compatible", "4-bit", "awq", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2_5_vl (Qwen2_5_VLForConditionalGeneration)", "total_params_b": 3.754622976, "active_params_b": 3.085938688, "layers": 36, "attention_heads": 16, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 128000, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "AWQ 4-bit + BF16/F16 side tensors", "weight_bytes_per_param": 0.5, "resident_weight_gb": 3.401637888, "active_weight_gb": 2.064269312, "kv_alloc_gb_per_1k_tokens": 0.036864, "kv_read_gb_per_1k_tokens": 0.036864, "notes": "Audited profile-backed row from live Hugging Face metadata on 2026-07-06; downloads=176189; data quality: API metadata, model card, served AWQ config, BF16 base config comparison, and direct safetensors header grouping. Resident weight uses exact stored tensor bytes; active weight excludes the resident-only visual tower while keeping tied model.embed_tokens.weight in swept decode traffic." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct-AWQ" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-VL-3B-Instruct-AWQ" }, { "label": "Pinned model config", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct-AWQ/raw/e7b623934290c5a4da0ee3c6e1e57bfb6b5abbf2/config.json" }, { "label": "Pinned safetensors artifact", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct-AWQ/resolve/e7b623934290c5a4da0ee3c6e1e57bfb6b5abbf2/model.safetensors" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct/raw/66285546d2b821cf421d4f5eb2576359d3770cd3/config.json" } ] }, { "id": "unsloth/Qwen3.5-0.8B-GGUF", "name": "Qwen3.5-0.8B-GGUF GGUF", "short_name": "Qwen3.5 0.8B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 283700, "tags": [ "transformers", "gguf", "unsloth", "image-text-to-text", "base_model:Qwen/Qwen3.5-0.8B", "base_model:quantized:Qwen/Qwen3.5-0.8B", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 0.752393024, "active_params_b": 0.752393024, "layers": 24, "attention_heads": 8, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "hybrid", "weight_precision": "GGUF BF16/F32", "weight_bytes_per_param": 2.00132509468881, "resident_weight_gb": 1.516744736, "active_weight_gb": 1.50578304, "kv_alloc_gb_per_1k_tokens": 0.012288, "kv_read_gb_per_1k_tokens": 0.012288, "kv_alloc_fixed_gb": 0.019759104, "kv_read_fixed_gb": 0.019759104, "notes": "Audited on 2026-07-06 from live HF metadata, pinned Qwen base config, linked-object HEAD checks, and a direct GGUF header range-read. The selected API artifact is Qwen3.5-0.8B-BF16.gguf: linked file 1.516744736 GB, tensor spans 1.505783040 GB, and GGUF/header/tokenizer overhead 0.010961696 GB. Text decode uses 6 full-context attention layers plus 18 fixed-state DeltaNet layers; separate mmproj sidecars are not included." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/Qwen3.5-0.8B-GGUF" }, { "label": "Pinned base model config", "url": "https://huggingface.co/Qwen/Qwen3.5-0.8B/raw/2fc06364715b967f1860aea9cf38778875588b17/config.json" }, { "label": "Pinned selected BF16 GGUF artifact", "url": "https://huggingface.co/unsloth/Qwen3.5-0.8B-GGUF/resolve/6ab461498e2023f6e3c1baea90a8f0fe38ab64d0/Qwen3.5-0.8B-BF16.gguf" }, { "label": "HF API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3.5-0.8B-GGUF" } ] }, { "id": "nvidia/DeepSeek-V4-Flash-NVFP4", "name": "DeepSeek-V4-Flash-NVFP4 NVFP4", "short_name": "DeepSeek V4 Flash NVFP4", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 449002, "tags": [ "Model Optimizer", "safetensors", "deepseek_v4", "nvidia", "ModelOpt", "DeepSeekV4", "quantized", "NVFP4", "nvfp4", "text-generation", "base_model:deepseek-ai/DeepSeek-V4-Flash", "base_model:quantized:deepseek-ai/DeepSeek-V4-Flash", "license:mit", "8-bit", "fp8", "region:us" ], "architecture": { "type": "ModelOpt NVFP4 MoE transformer", "detail": "deepseek_v4 CSA/HCA MoE (DeepseekV4ForCausalLM)", "total_params_b": 284, "active_params_b": 13, "layers": 43, "attention_heads": 64, "kv_heads": 1, "head_dim": 512, "max_context_tokens": 1048576, "routed_experts": 256, "routed_experts_per_token": 6, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "ModelOpt NVFP4 experts", "weight_bytes_per_param": 0.5, "resident_weight_gb": 168.266793544, "active_weight_gb": 11.439094028, "kv_alloc_gb_per_1k_tokens": 0.00688, "kv_read_gb_per_1k_tokens": 0.001504, "kv_alloc_fixed_gb": 0.017842176, "kv_read_fixed_gb": 0.016646144, "expert_param_b": 0.6086994, "shared_expert_param_b": 0, "fixed_param_b": 7.786897628, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, NVIDIA model card, pinned served config, hf_quant_config, official DeepSeek V4 Flash config/code comparison, linked-object metadata, safetensors index, and direct range-read shard headers. Catalog parameters use the model-card logical 284B total / 13B active values. Bound weights use NVIDIA-specific stored bytes: 168.266793544 GB resident tensor payload, 7.786897628 GB fixed ordinary traffic, 0.608699400 GB per routed expert index, and 4.652849516 GB resident-only embed/MTP bytes. KV/state traffic reuses the audited CSA/HCA BF16 path because the ModelOpt sidecar records kv_cache_quant_algo null." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/DeepSeek-V4-Flash-NVFP4" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/DeepSeek-V4-Flash-NVFP4" }, { "label": "Pinned model config", "url": "https://huggingface.co/nvidia/DeepSeek-V4-Flash-NVFP4/raw/e3cd60e7de98e9867116860d522499a728de1cf9/config.json" }, { "label": "Pinned ModelOpt quantization config", "url": "https://huggingface.co/nvidia/DeepSeek-V4-Flash-NVFP4/raw/e3cd60e7de98e9867116860d522499a728de1cf9/hf_quant_config.json" }, { "label": "Safetensors index and shard-header audit", "url": "https://huggingface.co/nvidia/DeepSeek-V4-Flash-NVFP4/resolve/e3cd60e7de98e9867116860d522499a728de1cf9/model.safetensors.index.json" }, { "label": "Official DeepSeek V4 Flash inference review", "url": "https://huggingface.co/nvidia/DeepSeek-V4-Flash-NVFP4/raw/e3cd60e7de98e9867116860d522499a728de1cf9/inference/model.py" } ] }, { "id": "QuantTrio/Qwen3-Coder-30B-A3B-Instruct-AWQ", "name": "Qwen3-Coder-30B-A3B-Instruct-AWQ 4-bit", "short_name": "Qwen3 Coder 30B A3B Instruct", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 267003, "tags": [ "transformers", "safetensors", "qwen3_moe", "text-generation", "AWQ", "\u91cf\u5316\u4fee\u590d", "vLLM", "conversational", "arxiv:2505.09388", "base_model:Qwen/Qwen3-Coder-30B-A3B-Instruct", "base_model:quantized:Qwen/Qwen3-Coder-30B-A3B-Instruct", "license:apache-2.0", "endpoints_compatible", "4-bit", "awq", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_moe (Qwen3MoeForCausalLM)", "total_params_b": 30.532122624, "active_params_b": 3.3, "layers": 48, "attention_heads": 32, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "AWQ 4-bit + BF16/F16 side tensors", "weight_bytes_per_param": 0.5, "resident_weight_gb": 16.80267264, "active_weight_gb": 2.059956224, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "expert_param_b": 0.235339776, "shared_expert_param_b": 0, "fixed_param_b": 2.23719424, "notes": "Manually audited on 2026-07-06 from current 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. Bound weights use exact stored AWQ tensor payload bytes: 16.802672640 GB resident package, 1.118597120 GB fixed ordinary traffic, 0.117669888 GB per routed expert index, and 0.622329856 GB resident-only input embedding. The repo metadata identifies the official Qwen3-Coder base, but served config differences mean the audited profile is authoritative for this package." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/QuantTrio/Qwen3-Coder-30B-A3B-Instruct-AWQ" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/QuantTrio/Qwen3-Coder-30B-A3B-Instruct-AWQ" }, { "label": "Pinned served AWQ config", "url": "https://huggingface.co/QuantTrio/Qwen3-Coder-30B-A3B-Instruct-AWQ/raw/c58857a7f41c0920f73d1b56678640f9c02017d7/config.json" }, { "label": "Official BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct/raw/b2cff646eb4bb1d68355c01b18ae02e7cf42d120/config.json" }, { "label": "Safetensors index and shard-header audit", "url": "https://huggingface.co/QuantTrio/Qwen3-Coder-30B-A3B-Instruct-AWQ/resolve/c58857a7f41c0920f73d1b56678640f9c02017d7/model.safetensors.index.json" } ] }, { "id": "stepfun-ai/Step-3.7-Flash-NVFP4", "name": "Step-3.7-Flash-NVFP4 NVFP4", "short_name": "Step 3.7 Flash NVFP4", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 241770, "tags": [ "transformers", "safetensors", "step3p7", "text-generation", "vision-language", "multimodal", "moe", "image-text-to-text", "conversational", "custom_code", "en", "license:apache-2.0", "8-bit", "modelopt", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "step3p7 VLM with step3p5 MoE language model", "total_params_b": 198, "active_params_b": 11, "layers": 45, "attention_heads": 64, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 288, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "NVFP4 + BF16/F8 side tensors", "weight_bytes_per_param": 0.5, "resident_weight_gb": 129.241354856, "active_weight_gb": 15.3218314, "kv_alloc_gb_per_1k_tokens": 0.024576, "kv_read_gb_per_1k_tokens": 0.024576, "kv_alloc_fixed_gb": 0.034603008, "kv_read_fixed_gb": 0.034603008, "expert_param_b": 0.743180256, "shared_expert_param_b": 0, "fixed_param_b": 24.698220752, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, model card, pinned served config, ModelOpt quantization sidecar, BF16 config comparison, remote modeling code, safetensors index, and direct shard-header range reads. Catalog parameters use the model-card 198B total / roughly 11B active description. Bound weights use exact stored bytes: 129.241354856 GB resident package, 12.349110376 GB fixed ordinary traffic, 0.371590128 GB per routed expert index across 288 experts, and 9.874287616 GB resident-only input embedding, vision/projector, and MTP tensors. KV traffic uses FP8 full/sliding cache: 12 full-context layers plus 33 sliding-window layers capped at 512 tokens." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/stepfun-ai/Step-3.7-Flash-NVFP4" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/stepfun-ai/Step-3.7-Flash-NVFP4" }, { "label": "Pinned served config", "url": "https://huggingface.co/stepfun-ai/Step-3.7-Flash-NVFP4/raw/4275532ffd9a9496ff36b7a2dc4a9db1048da438/config.json" }, { "label": "Pinned ModelOpt quantization config", "url": "https://huggingface.co/stepfun-ai/Step-3.7-Flash-NVFP4/raw/4275532ffd9a9496ff36b7a2dc4a9db1048da438/hf_quant_config.json" }, { "label": "BF16 Step 3.7 Flash config", "url": "https://huggingface.co/stepfun-ai/Step-3.7-Flash/raw/5f6244077ac62e04eec3f320501ff8c2b293373a/config.json" }, { "label": "Safetensors index and shard-header audit", "url": "https://huggingface.co/stepfun-ai/Step-3.7-Flash-NVFP4/resolve/4275532ffd9a9496ff36b7a2dc4a9db1048da438/model.safetensors.index.json" } ] }, { "id": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B", "name": "DeepSeek-R1-Distill-Llama-8B BF16", "short_name": "DeepSeek R1 Distill Llama 8B", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 305879, "tags": [ "transformers", "safetensors", "llama", "text-generation", "conversational", "arxiv:2501.12948", "license:mit", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "llama (LlamaForCausalLM)", "total_params_b": 8.030261248, "active_params_b": 7.504924672, "layers": 32, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 16.060522496, "active_weight_gb": 15.009849344, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "notes": "Manually corrected on 2026-07-06 from current Hugging Face API metadata, pinned DeepSeek model card, pinned served config, generation config, safetensors index, and direct range-read safetensors shard headers. Legacy fallback values mirror the audited profile: exact 16.060522496 GB resident BF16 package, 15.009849344 GB ordinary text-decode swept traffic, 1.050673152 GB resident-only input embedding, and BF16 full-context KV. The authoritative profile supersedes this legacy adapter data for production bounds." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-R1-Distill-Llama-8B" }, { "label": "Hugging Face model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B" }, { "label": "Model config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B/raw/6a6f4aa4197940add57724a7707d069478df56b1/config.json" }, { "label": "Generation config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B/raw/6a6f4aa4197940add57724a7707d069478df56b1/generation_config.json" }, { "label": "Safetensors index", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B/raw/6a6f4aa4197940add57724a7707d069478df56b1/model.safetensors.index.json" } ] }, { "id": "google/gemma-4-26B-A4B-it-assistant", "name": "Gemma 4 26B A4B IT Assistant Drafter BF16", "short_name": "Gemma 4 26B A4B IT Assistant", "license": "apache-2.0", "hf_pipeline_tag": "any-to-any", "hf_downloads": 318820, "tags": [ "transformers", "safetensors", "gemma4_assistant", "text-generation", "any-to-any", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_assistant", "total_params_b": 0.419711236, "active_params_b": 0.419711236, "layers": 4, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 0.839422472, "active_weight_gb": 0.839422472, "kv_alloc_gb_per_1k_tokens": 0.026624, "kv_read_gb_per_1k_tokens": 0.026624, "notes": "Manually reviewed on 2026-07-06 from current Hugging Face API metadata, served config, generation config, model card, and direct safetensors header range read. This repo is a Gemma 4 MTP speculative drafter used as assistant_model with google/gemma-4-26B-A4B-it. The profile records the exact 0.839422472 GB resident BF16 drafter artifact and intentionally disables production throughput because Bounds Engine v1 lacks paired target+assistant speculative decoding." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it-assistant" }, { "label": "Model config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it-assistant/raw/44033eb52554fc5339b383c81309119c684e4fb0/config.json" }, { "label": "Generation config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it-assistant/raw/44033eb52554fc5339b383c81309119c684e4fb0/generation_config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-26B-A4B-it-assistant" }, { "label": "Target model card", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it-assistant/resolve/44033eb52554fc5339b383c81309119c684e4fb0/model.safetensors" } ] }, { "id": "ISTA-DASLab/gemma-3-27b-it-GPTQ-4b-128g", "name": "Gemma 3 27B IT GPTQ 4B 128G INT4", "short_name": "Gemma 3 27B IT GPTQ 4B", "license": "gemma", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 236301, "tags": [ "transformers", "safetensors", "gemma3", "image-text-to-text", "int4", "vllm", "llmcompressor", "conversational", "base_model:google/gemma-3-27b-it", "base_model:quantized:google/gemma-3-27b-it", "license:gemma", "text-generation-inference", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense multimodal transformer", "detail": "gemma3_text (Gemma3ForConditionalGeneration)", "total_params_b": 27.632394708, "active_params_b": 27.209334372, "layers": 62, "attention_heads": 32, "kv_heads": 16, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "compressed-tensors GPTQ INT4", "weight_bytes_per_param": 0.6104158688467444, "resident_weight_gb": 16.867252224, "active_weight_gb": 16.021131552, "kv_alloc_gb_per_1k_tokens": 0.08192, "kv_read_gb_per_1k_tokens": 0.08192, "notes": "Manually audited on 2026-07-06 from current HF API metadata, served compressed-tensors config, generation config, model card metadata, safetensors index, direct safetensors header range reads, and Gemma 3 local/global attention documentation. downloads=236301. Exact tensor payload bytes: 16.867252224 GB resident, 16.021131552 GB swept ordinary text traffic, and 0.846120672 GB resident-only vision/projector bytes. The package has no separate lm_head tensor, so the BF16 input embedding is charged as tied output-projection traffic. KV fields show only the full-context component per 1K tokens; the audited profile also charges 52 sliding-window layers capped at 1024 tokens." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/ISTA-DASLab/gemma-3-27b-it-GPTQ-4b-128g" }, { "label": "Model config", "url": "https://huggingface.co/ISTA-DASLab/gemma-3-27b-it-GPTQ-4b-128g/raw/9c5f61a23e1ccf3921173c2b32f6ff0af1bffd1b/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-3-27b-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/ISTA-DASLab/gemma-3-27b-it-GPTQ-4b-128g" }, { "label": "Generation config", "url": "https://huggingface.co/ISTA-DASLab/gemma-3-27b-it-GPTQ-4b-128g/raw/9c5f61a23e1ccf3921173c2b32f6ff0af1bffd1b/generation_config.json" }, { "label": "Safetensors index", "url": "https://huggingface.co/ISTA-DASLab/gemma-3-27b-it-GPTQ-4b-128g/raw/9c5f61a23e1ccf3921173c2b32f6ff0af1bffd1b/model.safetensors.index.json" } ] }, { "id": "XiaomiMiMo/MiMo-7B-RL", "name": "MiMo-7B-RL BF16", "short_name": "MiMo 7B RL", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 405357, "tags": [ "transformers", "safetensors", "mimo", "text-generation", "conversational", "custom_code", "arxiv:2505.07608", "license:mit", "eval-results", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "mimo (MiMoForCausalLM)", "total_params_b": 7.833409536, "active_params_b": 7.001337856, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 15.666819072, "active_weight_gb": 14.002675712, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, pinned model card, served config, generation config, custom configuration/modeling code, safetensors index, and direct safetensors shard header range reads. downloads=405357. Exact tensor payload bytes: 15.666819072 GB resident, 14.002675712 GB ordinary text-decode swept traffic, and 1.664143360 GB resident-only input embedding plus MTP tensors. This catalog row models ordinary target-model decode; MiMo-MTP speculative serving requires a separate adapter." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/XiaomiMiMo/MiMo-7B-RL" }, { "label": "Model config", "url": "https://huggingface.co/XiaomiMiMo/MiMo-7B-RL/raw/6299b5a2c45daf0c429285c92b8e61a5bd011c0d/config.json" }, { "label": "Generation config", "url": "https://huggingface.co/XiaomiMiMo/MiMo-7B-RL/raw/6299b5a2c45daf0c429285c92b8e61a5bd011c0d/generation_config.json" }, { "label": "Custom modeling code", "url": "https://huggingface.co/XiaomiMiMo/MiMo-7B-RL/raw/6299b5a2c45daf0c429285c92b8e61a5bd011c0d/modeling_mimo.py" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/XiaomiMiMo/MiMo-7B-RL" }, { "label": "Safetensors index", "url": "https://huggingface.co/XiaomiMiMo/MiMo-7B-RL/raw/6299b5a2c45daf0c429285c92b8e61a5bd011c0d/model.safetensors.index.json" } ] }, { "id": "nvidia/Llama-3.1-8B-Instruct-NVFP4", "name": "Llama-3.1-8B-Instruct-NVFP4 NVFP4", "short_name": "Llama 3.1 8B Instruct NVFP4", "license": "other", "hf_pipeline_tag": "unknown", "hf_downloads": 280705, "tags": [ "Model Optimizer", "safetensors", "llama", "nvidia", "ModelOpt", "llama3", "quantized", "FP4", "fp4", "base_model:meta-llama/Llama-3.1-8B-Instruct", "base_model:quantized:meta-llama/Llama-3.1-8B-Instruct", "license:other", "8-bit", "modelopt", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "llama (LlamaForCausalLM)", "total_params_b": 8.0303, "active_params_b": 8.0303, "layers": 32, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "ModelOpt NVFP4 + FP8 scales + BF16 fixed tensors", "weight_bytes_per_param": 0.5, "resident_weight_gb": 6.027749376, "active_weight_gb": 4.977076224, "kv_alloc_gb_per_1k_tokens": 0.065536, "kv_read_gb_per_1k_tokens": 0.065536, "notes": "Manually corrected on 2026-07-06 from current Hugging Face API metadata, NVIDIA model card, pinned served config, ModelOpt quantization config, NVIDIA FP8 sibling config comparison, safetensors index, linked-object HEAD checks, and direct safetensors shard-header range reads. Legacy fallback values mirror the audited profile: exact 6.027749376 GB stored resident package, 4.977076224 GB ordinary text-decode swept traffic, 1.050673152 GB resident-only input embedding, and FP8 full-context KV. The authoritative profile supersedes the generated simple 0.5-byte NVFP4 resident estimate and BF16 KV estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/Llama-3.1-8B-Instruct-NVFP4" }, { "label": "Model config", "url": "https://huggingface.co/nvidia/Llama-3.1-8B-Instruct-NVFP4/raw/bdb54e24298451af785c0ac63c1b485e9b7400a2/config.json" }, { "label": "ModelOpt quantization config", "url": "https://huggingface.co/nvidia/Llama-3.1-8B-Instruct-NVFP4/raw/bdb54e24298451af785c0ac63c1b485e9b7400a2/hf_quant_config.json" }, { "label": "Safetensors index", "url": "https://huggingface.co/nvidia/Llama-3.1-8B-Instruct-NVFP4/raw/bdb54e24298451af785c0ac63c1b485e9b7400a2/model.safetensors.index.json" } ] }, { "id": "google/gemma-4-E4B-it-qat-q4_0-gguf", "name": "gemma-4-E4B-it-qat-q4_0-gguf GGUF 4-bit", "short_name": "gemma 4 E4B it qat q4 0", "license": "apache-2.0", "hf_pipeline_tag": "any-to-any", "hf_downloads": 301981, "tags": [ "gguf", "any-to-any", "base_model:google/gemma-4-E4B-it-qat-q4_0-unquantized", "base_model:quantized:google/gemma-4-E4B-it-qat-q4_0-unquantized", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 7.9411, "active_params_b": 7.9411, "layers": 42, "attention_heads": 8, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q4_0", "weight_bytes_per_param": 0.6907315678596464, "resident_weight_gb": 5.154939136, "active_weight_gb": 2.82700832, "kv_alloc_gb_per_1k_tokens": 0.086016, "kv_read_gb_per_1k_tokens": 0.086016, "notes": "Manually corrected on 2026-07-06 from current Hugging Face API metadata, pinned official Google GGUF model card, pinned QAT unquantized base config, QAT release model card, GGUF linked-object HEAD checks, and a direct selected-file GGUF header range read. Bound weights use the selected Q4_0 main file size 5.154939136 GB for residency and 2.82700832 GB of ordinary swept tensor traffic after excluding the resident per-layer token embedding lookup table. The separate 0.991551904 GB mmproj sidecar is not included unless explicitly loaded." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-4-E4B-it-qat-q4_0-gguf" }, { "label": "Base model config", "url": "https://huggingface.co/google/gemma-4-E4B-it-qat-q4_0-unquantized/raw/dfc5b925ddb1d41aaf1fe9679abdcfb0805e1aa6/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-E4B-it-qat-q4_0-gguf" }, { "label": "Selected GGUF header audit", "url": "https://huggingface.co/google/gemma-4-E4B-it-qat-q4_0-gguf/resolve/bb3b92e6f031fa438b409f898dd9f14f499a0cb0/gemma-4-E4B_q4_0-it.gguf" } ] }, { "id": "meta-llama/Llama-2-7b-chat-hf", "name": "Llama-2-7b-chat-hf F16", "short_name": "Llama 2 7b chat", "license": "llama2", "hf_pipeline_tag": "text-generation", "hf_downloads": 265563, "tags": [ "transformers", "pytorch", "safetensors", "llama", "text-generation", "facebook", "meta", "llama-2", "conversational", "en", "arxiv:2307.09288", "license:llama2", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "unknown", "total_params_b": 6.7384, "active_params_b": 6.7384, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "F16", "weight_bytes_per_param": 2, "resident_weight_gb": 13.476835328, "active_weight_gb": 13.476835328, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manually corrected on 2026-07-06 from current Hugging Face API metadata and authenticated access checks. The API records 6738417664 F16 safetensors parameters / 13.476835328 GB and gated: manual. Raw config/index requests and authenticated HF CLI downloads as osolmaz were denied, so the authoritative profile fails closed instead of inferring KV geometry or context length from the model name." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/meta-llama/Llama-2-7b-chat-hf" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-2-7b-chat-hf" }, { "label": "Gated config access check", "url": "https://huggingface.co/meta-llama/Llama-2-7b-chat-hf/raw/f5db02db724555f92da89c216ac04704f23d4590/config.json" } ] }, { "id": "unsloth/Qwen3-Coder-Next-GGUF", "name": "Qwen3-Coder-Next-GGUF GGUF", "short_name": "Qwen3 Coder Next", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 233649, "tags": [ "gguf", "qwen3_next", "unsloth", "qwen", "qwen3", "text-generation", "base_model:Qwen/Qwen3-Coder-Next", "base_model:quantized:Qwen/Qwen3-Coder-Next", "license:apache-2.0", "endpoints_compatible", "region:us", "imatrix", "conversational" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_next (Qwen3NextForCausalLM)", "total_params_b": 79.674391296, "active_params_b": 3.56376448, "layers": 48, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "GGUF IQ4_NL", "weight_bytes_per_param": 0.5664347570894557, "resident_weight_gb": 45.13034448, "active_weight_gb": 2.312127488, "expert_param_b": 0.14994605281007936, "fixed_param_b": 2.5824349754176303, "shared_expert_param_b": 0, "kv_alloc_gb_per_1k_tokens": 0.024576, "kv_read_gb_per_1k_tokens": 0.024576, "kv_alloc_fixed_gb": 0.077856768, "kv_read_fixed_gb": 0.077856768, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, model card, Qwen base config, selected linked-object HEAD check, direct GGUF header/tensor-index range read, and Qwen3-Next runtime review. The audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/Qwen3-Coder-Next-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3-Coder-Next-GGUF" }, { "label": "Selected IQ4_NL GGUF artifact", "url": "https://huggingface.co/unsloth/Qwen3-Coder-Next-GGUF/resolve/ce09c67b53bc8739eef83fe67b2f5d293c270632/Qwen3-Coder-Next-IQ4_NL.gguf" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3-Coder-Next/raw/a7fbcb5c0e12d62a448eaa0e260346bf5dcc0feb/config.json" }, { "label": "Transformers Qwen3-Next implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/qwen3_next/modeling_qwen3_next.py" } ] }, { "id": "Jackrong/Qwopus3.6-27B-Coder-MTP-GGUF", "name": "Qwopus3.6-27B-Coder-MTP-GGUF GGUF", "short_name": "Qwopus3.6 27B Coder MTP", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 297857, "tags": [ "transformers", "gguf", "text-generation-inference", "unsloth", "qwen3_6", "reasoning", "chain-of-thought", "lora", "sft", "agent", "tool-use", "function-calling", "coder", "text-generation", "en", "zh", "es", "ru", "ja", "dataset:Jackrong/Claude-opus-4.6-TraceInversion-9000x", "dataset:Jackrong/Claude-opus-4.7-TraceInversion-5000x", "dataset:lambda/hermes-agent-reasoning-traces", "base_model:Jackrong/Qwopus3.6-27B-v2", "base_model:adapter:Jackrong/Qwopus3.6-27B-v2", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 27.320697856, "active_params_b": 25.624600064, "layers": 64, "attention_heads": 24, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q5_K_M", "weight_bytes_per_param": 0.715051225959431, "resident_weight_gb": 19.535698496, "active_weight_gb": 18.346018816, "kv_alloc_gb_per_1k_tokens": 0.065536, "kv_read_gb_per_1k_tokens": 0.065536, "kv_alloc_fixed_gb": 0.154927104, "kv_read_fixed_gb": 0.154927104, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, model card, Jackrong base config, Qwen3.6 base config comparison, selected linked-object HEAD checks, direct GGUF header/tensor-index range read, and Qwen3.5 runtime review. The API GGUF block points at mmproj-F32, so the audited profile uses the card-benchmark Q5_K_M main artifact and supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Jackrong/Qwopus3.6-27B-Coder-MTP-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Jackrong/Qwopus3.6-27B-Coder-MTP-GGUF" }, { "label": "Selected Q5_K_M GGUF artifact", "url": "https://huggingface.co/Jackrong/Qwopus3.6-27B-Coder-MTP-GGUF/resolve/ab9175af03e34068fd773eeb64c80461bec2a526/Qwopus3.6-27B-Coder-MTP-Q5_K_M.gguf" }, { "label": "Base model config", "url": "https://huggingface.co/Jackrong/Qwopus3.6-27B-v2/raw/d0d82f4ccc9d41d4fe9595e96be4595327bb5de7/config.json" }, { "label": "Qwen3.6 base config", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json" }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py" } ] }, { "id": "bartowski/gemma-2-2b-it-GGUF", "name": "gemma-2-2b-it-GGUF GGUF F32", "short_name": "gemma 2 2b it", "license": "gemma", "hf_pipeline_tag": "text-generation", "hf_downloads": 258944, "tags": [ "transformers", "gguf", "conversational", "text-generation", "base_model:google/gemma-2-2b-it", "base_model:quantized:google/gemma-2-2b-it", "license:gemma", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma2", "total_params_b": 2.614341888, "active_params_b": 2.614341888, "layers": 26, "attention_heads": 8, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 8192, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF F32", "weight_bytes_per_param": 4.002312744185354, "resident_weight_gb": 10.463413856, "active_weight_gb": 10.457367552, "kv_alloc_gb_per_1k_tokens": 0.053248, "kv_read_gb_per_1k_tokens": 0.053248, "kv_alloc_fixed_gb": 0.218103808, "kv_read_fixed_gb": 0.218103808, "notes": "Audited on 2026-07-06 from live Hugging Face metadata and direct GGUF header range-read at commit 855f67caed130e1befc571b52bd181be2e858883. The API-selected artifact is gemma-2-2b-it-f32.gguf. Legacy adapter fields approximate the Gemma 2 layered KV rule as 13 full-context layers plus 13 sliding-window layers capped at 4096 tokens; the self-contained model profile is authoritative for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/bartowski/gemma-2-2b-it-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/google/gemma-2-2b-it/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-2-2b-it" } ] }, { "id": "google/gemma-4-E2B", "name": "gemma-4-E2B BF16", "short_name": "gemma 4 E2B", "license": "apache-2.0", "hf_pipeline_tag": "any-to-any", "hf_downloads": 184194, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "any-to-any", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text + multimodal PLE", "total_params_b": 5.123178979, "active_params_b": 2.298639651, "layers": 35, "attention_heads": 8, "kv_heads": 1, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 10.246357958, "active_weight_gb": 4.597279302, "kv_alloc_gb_per_1k_tokens": 0.006144, "kv_read_gb_per_1k_tokens": 0.014336, "kv_alloc_fixed_gb": 0.006291456, "kv_read_fixed_gb": 0.014680064, "notes": "Audited on 2026-07-06 from live Hugging Face metadata, pinned config comparison, and direct safetensors header range-read at commit 63db66a33dc06d58c02b1e887446e103c202602c. Active weight is ordinary text-decode swept traffic excluding the resident-only per-layer embedding table and multimodal towers. Legacy adapter fields approximate the shared-KV rule: full-context traffic is token-proportional, while the sliding-window component is capped at 512 tokens; the self-contained model profile is authoritative for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-4-E2B" }, { "label": "Model config", "url": "https://huggingface.co/google/gemma-4-E2B/raw/63db66a33dc06d58c02b1e887446e103c202602c/config.json" }, { "label": "Instruction-tuned config comparison", "url": "https://huggingface.co/google/gemma-4-E2B-it/raw/70af34e20bd4b7a91f0de6b22675850c43922a03/config.json" }, { "label": "Safetensors header", "url": "https://huggingface.co/google/gemma-4-E2B/resolve/63db66a33dc06d58c02b1e887446e103c202602c/model.safetensors" } ] }, { "id": "sugoitoolkit/Sugoi-32B-Ultra-GGUF", "name": "Sugoi-32B-Ultra-GGUF GGUF", "short_name": "Sugoi 32B Ultra", "license": "apache-2.0", "hf_pipeline_tag": "translation", "hf_downloads": 251014, "tags": [ "gguf", "translation", "ja", "en", "base_model:Qwen/Qwen2.5-32B-Instruct", "base_model:quantized:Qwen/Qwen2.5-32B-Instruct", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 32.7639, "active_params_b": 32.7639, "layers": 64, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF F16", "weight_bytes_per_param": 2.000250781559292, "resident_weight_gb": 65.53596928, "active_weight_gb": 63.972855808, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Audited on 2026-07-06 from live Hugging Face metadata, audited Qwen2.5 32B Instruct config/profile evidence, linked-object HEAD checks, and direct GGUF header range-read at commit 3167b2d18c29a12f6eece6f6941e8eeb52d29e08. The API-selected artifact is Sugoi-32B-Ultra-F16.gguf. Active weight excludes token_embd.weight input lookup and includes the separate output.weight projection." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/sugoitoolkit/Sugoi-32B-Ultra-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct/raw/5ede1c97bbab6ce5cda5812749b4c0bdf79b18dd/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct" }, { "label": "Selected F16 GGUF", "url": "https://huggingface.co/sugoitoolkit/Sugoi-32B-Ultra-GGUF/resolve/3167b2d18c29a12f6eece6f6941e8eeb52d29e08/Sugoi-32B-Ultra-F16.gguf" } ] }, { "id": "google/gemma-2-2b", "name": "gemma-2-2b F32", "short_name": "gemma 2 2b", "license": "gemma", "hf_pipeline_tag": "text-generation", "hf_downloads": 270278, "tags": [ "transformers", "safetensors", "gemma2", "text-generation", "arxiv:2009.03300", "arxiv:1905.07830", "arxiv:1911.11641", "arxiv:1904.09728", "arxiv:1905.10044", "arxiv:1907.10641", "arxiv:1811.00937", "arxiv:1809.02789", "arxiv:1911.01547", "arxiv:1705.03551", "arxiv:2107.03374", "arxiv:2108.07732", "arxiv:2110.14168", "arxiv:2009.11462", "arxiv:2101.11718", "arxiv:2110.08193", "arxiv:1804.09301", "arxiv:2109.07958", "arxiv:1804.06876", "arxiv:2103.03874", "arxiv:2304.06364", "arxiv:1903.00161", "arxiv:2206.04615", "arxiv:2203.09509", "arxiv:2403.13793", "license:gemma", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma2 (gated config unavailable)", "total_params_b": 2.614341888, "active_params_b": 2.614341888, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "F32", "weight_bytes_per_param": 4, "resident_weight_gb": 10.457367552, "active_weight_gb": 10.457367552, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manual fail-closed correction on 2026-07-06 from current Hugging Face API metadata and direct hf download access checks. The repo is manually gated; config and tensor headers are inaccessible, so production bounds use the unsupported profile and must not infer KV geometry or swept traffic from this legacy row." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-2-2b" } ] }, { "id": "unsloth/diffusiongemma-26B-A4B-it-GGUF", "name": "diffusiongemma-26B-A4B-it-GGUF GGUF", "short_name": "diffusiongemma 26B A4B it", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 286450, "tags": [ "gguf", "gemma4", "unsloth", "gemma", "google", "diffusion_gemma", "image-text-to-text", "base_model:google/diffusiongemma-26B-A4B-it", "base_model:quantized:google/diffusiongemma-26B-A4B-it", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Block diffusion MoE transformer", "detail": "diffusion-gemma GGUF block diffusion", "total_params_b": 25.250987068, "active_params_b": 3.8, "layers": 30, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "GGUF BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 50.5408344, "active_weight_gb": 7.703772992, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manual fail-closed correction on 2026-07-06 from current HF API commit f4183a2c7a354128d02545752303c4354d165bf0, model card, linked GGUF file sizes, and direct BF16 GGUF tensor-index range read. The exact selected BF16 GGUF resident package is 50.540834400 GB. Production bounds are disabled by the unsupported profile because DiffusionGemma uses block diffusion over a 256-token canvas and needs a dedicated llama-diffusion-cli runner, which Bounds Engine v1 does not model as ordinary autoregressive decode." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/diffusiongemma-26B-A4B-it-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/google/diffusiongemma-26B-A4B-it/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/diffusiongemma-26B-A4B-it" } ] }, { "id": "Qwen/Qwen3-14B-FP8", "name": "Qwen3-14B-FP8 FP8", "short_name": "Qwen3 14B FP8", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 287664, "tags": [ "transformers", "safetensors", "qwen3", "text-generation", "conversational", "arxiv:2309.00071", "arxiv:2505.09388", "base_model:Qwen/Qwen3-14B", "base_model:quantized:Qwen/Qwen3-14B", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "fp8", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 14.7691136, "active_params_b": 13.99120128, "layers": 40, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "FP8", "weight_bytes_per_param": 1.1054265030502575, "resident_weight_gb": 16.3261696, "active_weight_gb": 14.77034496, "kv_alloc_gb_per_1k_tokens": 0.16384, "kv_read_gb_per_1k_tokens": 0.16384, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, pinned served FP8 config, audited BF16 base config/profile comparison, safetensors index, and direct shard header range reads. Bound weights use exact direct tensor payload bytes: 16.326169600 GB resident and 14.770344960 GB ordinary text-decode swept traffic after excluding only model.embed_tokens.weight. The audited profile supersedes the generated rounded FP8 estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-14B-FP8" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-14B-FP8/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-14B" } ] }, { "id": "MaziyarPanahi/Qwen3-0.6B-GGUF", "name": "Qwen3-0.6B-GGUF GGUF F16", "short_name": "Qwen3 0.6B", "license": "unknown", "hf_pipeline_tag": "text-generation", "hf_downloads": 231812, "tags": [ "gguf", "mistral", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:Qwen/Qwen3-0.6B", "base_model:quantized:Qwen/Qwen3-0.6B", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 GGUF (Qwen3ForCausalLM)", "total_params_b": 0.751632384, "active_params_b": 0.59604992, "layers": 28, "attention_heads": 16, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF F16", "weight_bytes_per_param": 2, "resident_weight_gb": 1.509347456, "active_weight_gb": 1.192230912, "kv_alloc_gb_per_1k_tokens": 0.114688, "kv_read_gb_per_1k_tokens": 0.114688, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, model card, audited BF16 base profile comparison, linked GGUF file sizes, and direct selected F16 GGUF tensor-index range read. The authoritative bounds profile uses Qwen3-0.6B.fp16.gguf because it matches API gguf.totalFileSize: 1.509347456 GB resident and 1.192230912 GB ordinary text-decode swept traffic after excluding token_embd.weight." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-0.6B-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-0.6B-GGUF/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-0.6B" } ] }, { "id": "google/gemma-4-31B-it-qat-q4_0-gguf", "name": "gemma-4-31B-it-qat-q4_0-gguf GGUF 4-bit", "short_name": "gemma 4 31B it qat q4 0", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 284997, "tags": [ "transformers", "gguf", "image-text-to-text", "base_model:google/gemma-4-31B-it-qat-q4_0-unquantized", "base_model:quantized:google/gemma-4-31B-it-qat-q4_0-unquantized", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 30.697345596, "active_params_b": 30.697345596, "layers": 60, "attention_heads": 32, "kv_heads": 16, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q4_0 QAT", "weight_bytes_per_param": 0.5750008384536024, "resident_weight_gb": 17.650999456, "active_weight_gb": 17.635168128, "kv_alloc_gb_per_1k_tokens": 0.86016, "kv_read_gb_per_1k_tokens": 0.86016, "notes": "Manually audited on 2026-07-06 from HF API commit 4a311c5261daa0702f80836f8866114943651ab0, Google QAT unquantized base config, model card, linked GGUF file sizes, and a GGUF tensor-index range read. The authoritative bounds profile uses the selected gemma-4-31B_q4_0-it.gguf artifact: 17.650999456 GB resident, 17.635168128 GB swept text traffic, 0.015831328 GB resident-only GGUF metadata/header bytes, tied token_embd.weight output projection, and layered Gemma 4 hybrid KV. The separate mmproj GGUF file is not included unless explicitly loaded. Legacy per-1k KV fields show the short-context unsaturated coefficient; long-context calculation must use the profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-4-31B-it-qat-q4_0-gguf" }, { "label": "Base model config", "url": "https://huggingface.co/google/gemma-4-31B-it-qat-q4_0-unquantized/raw/4f926903562062220b3e54c1385c5ef2cd40bfd1/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-31B-it-qat-q4_0-unquantized" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-31B-it-qat-q4_0-gguf" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/google/gemma-4-31B-it-qat-q4_0-gguf/resolve/4a311c5261daa0702f80836f8866114943651ab0/gemma-4-31B_q4_0-it.gguf" } ] }, { "id": "Jackrong/Qwen3.5-9B-DeepSeek-V4-Flash-GGUF", "name": "Qwen3.5-9B-DeepSeek-V4-Flash-GGUF GGUF", "short_name": "Qwen3.5 9B DeepSeek V4 Flash", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 181436, "tags": [ "transformers", "gguf", "text-generation-inference", "unsloth", "qwen3_5", "reasoning", "distillation", "deepseek", "deepseek-v4", "sft", "long-cot", "chain-of-thought", "efficient-inference", "agent", "multilingual", "image-text-to-text", "en", "zh", "ko", "ja", "es", "ru", "dataset:Jackrong/DeepSeek-V4-Distill-8000x", "arxiv:2604.06628", "base_model:unsloth/Qwen3.5-9B", "base_model:quantized:unsloth/Qwen3.5-9B", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 8.953803264, "active_params_b": 7.936684544, "layers": 32, "attention_heads": 16, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q3_K_M", "weight_bytes_per_param": 0.516375096445273, "resident_weight_gb": 4.623521024, "active_weight_gb": 4.1755136, "kv_alloc_gb_per_1k_tokens": 0.032768, "kv_read_gb_per_1k_tokens": 0.032768, "kv_alloc_fixed_gb": 0.051904512, "kv_read_fixed_gb": 0.051904512, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, pinned model card, Unsloth base config, upstream Qwen config comparison, linked GGUF file sizes, and a direct GGUF tensor-index range read. The authoritative bounds profile uses the API-selected Qwen3.5-9B-DeepSeek-V4-Flash-Q3_K_M.gguf artifact: 4.623521024 GB resident, 4.175513600 GB swept text traffic, 0.448007424 GB resident-only token embedding plus GGUF metadata/header/file overhead, and Qwen3.5 hybrid full-attention KV plus fixed DeltaNet state. The separate mmproj GGUF sidecar is not included unless explicitly loaded." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Jackrong/Qwen3.5-9B-DeepSeek-V4-Flash-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/unsloth/Qwen3.5-9B/raw/005429cee5cb648998cf2b70eebdd83175989c9a/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/unsloth/Qwen3.5-9B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Jackrong/Qwen3.5-9B-DeepSeek-V4-Flash-GGUF" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/Jackrong/Qwen3.5-9B-DeepSeek-V4-Flash-GGUF/resolve/e842f7426b567ec1e335d5cc348f4c86cdcf1453/Qwen3.5-9B-DeepSeek-V4-Flash-Q3_K_M.gguf" } ] }, { "id": "Qwen/Qwen3-1.7B-GPTQ-Int8", "name": "Qwen3-1.7B-GPTQ-Int8 8-bit", "short_name": "Qwen3 1.7B GPTQ Int8", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 180833, "tags": [ "transformers", "safetensors", "qwen3", "text-generation", "conversational", "arxiv:2505.09388", "base_model:Qwen/Qwen3-1.7B", "base_model:quantized:Qwen/Qwen3-1.7B", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "8-bit", "gptq", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 1.720574976, "active_params_b": 1.720574976, "layers": 28, "attention_heads": 16, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GPTQ Int8", "weight_bytes_per_param": 1.2013183760264103, "resident_weight_gb": 2.066958336, "active_weight_gb": 2.066958336, "kv_alloc_gb_per_1k_tokens": 0.114688, "kv_read_gb_per_1k_tokens": 0.114688, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, pinned GPTQ config, model card, audited BF16 instruction profile comparison, linked-object HEAD check, and direct safetensors header range read. The authoritative bounds profile uses exact GPTQ tensor payload bytes: 2.066958336 GB resident and swept ordinary text traffic, tied token embedding/output projection with no separate lm_head.weight, and full-context FP16 KV. The safetensors linked object has 100120 bytes of header overhead outside tensor payload." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-1.7B-GPTQ-Int8" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-1.7B-GPTQ-Int8/raw/382fed3fa21d1a7e837b42a500564a0c4be4f060/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-1.7B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-1.7B-GPTQ-Int8" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/Qwen/Qwen3-1.7B-GPTQ-Int8/resolve/382fed3fa21d1a7e837b42a500564a0c4be4f060/model.safetensors" } ] }, { "id": "meta-llama/Llama-3.1-405B", "name": "Llama-3.1-405B BF16", "short_name": "Llama 3.1 405B", "license": "llama3.1", "hf_pipeline_tag": "text-generation", "hf_downloads": 230262, "tags": [ "transformers", "safetensors", "llama", "text-generation", "facebook", "meta", "pytorch", "llama-3", "en", "de", "fr", "it", "pt", "hi", "es", "th", "arxiv:2204.05149", "license:llama3.1", "eval-results", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Unknown gated model", "detail": "gated config unavailable", "total_params_b": 405.8533888, "active_params_b": 405.8533888, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 811.7067776, "active_weight_gb": 811.7067776, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Fail-closed on 2026-07-06 from current Hugging Face API metadata and access checks. The API records 405.853388800B BF16 safetensors parameters at commit b906e4dc842aa489c962f9db26554dcfdde901fe, or 811.706777600 GB if charged by dtype. Raw config, README, model.safetensors.index.json, a safetensors shard HEAD request, and hf download config.json all require manual approval in this environment. KV geometry, max context, tied embeddings, and swept decode traffic are intentionally unknown until direct gated evidence is available." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/meta-llama/Llama-3.1-405B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.1-405B" }, { "label": "Gated config access check", "url": "https://huggingface.co/meta-llama/Llama-3.1-405B/raw/b906e4dc842aa489c962f9db26554dcfdde901fe/config.json" }, { "label": "Gated tensor index access check", "url": "https://huggingface.co/meta-llama/Llama-3.1-405B/raw/b906e4dc842aa489c962f9db26554dcfdde901fe/model.safetensors.index.json" } ] }, { "id": "stelterlab/Mistral-Small-24B-Instruct-2501-AWQ", "name": "Mistral-Small-24B-Instruct-2501-AWQ 4-bit", "short_name": "Mistral Small 24B Instruct 2501", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 235342, "tags": [ "vllm", "safetensors", "mistral", "text-generation", "transformers", "conversational", "en", "fr", "de", "es", "it", "pt", "zh", "ja", "ru", "ko", "base_model:mistralai/Mistral-Small-24B-Instruct-2501", "base_model:quantized:mistralai/Mistral-Small-24B-Instruct-2501", "license:apache-2.0", "text-generation-inference", "4-bit", "awq", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "mistral (MistralForCausalLM)", "total_params_b": 23.5724032, "active_params_b": 22.90131456, "layers": 40, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "AWQ 4-bit + F16 side tensors", "weight_bytes_per_param": 0.6038529665061898, "resident_weight_gb": 14.2342656, "active_weight_gb": 12.89208832, "kv_alloc_gb_per_1k_tokens": 0.16384, "kv_read_gb_per_1k_tokens": 0.16384, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, model card, pinned AWQ config, current base config comparison, params.json, safetensors index, linked-object HEAD checks, and direct range-read shard headers. Exact tensor payload bytes: 14.234265600 GB resident, 12.892088320 GB swept ordinary text traffic, and 1.342177280 GB resident-only input embedding bytes. The safetensors linked objects have 105048 bytes of header overhead outside tensor payload." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/stelterlab/Mistral-Small-24B-Instruct-2501-AWQ" }, { "label": "Model config", "url": "https://huggingface.co/stelterlab/Mistral-Small-24B-Instruct-2501-AWQ/raw/cbda099649a0188dd888d44f0e4964d8d982dc9a/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/mistralai/Mistral-Small-24B-Instruct-2501" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/stelterlab/Mistral-Small-24B-Instruct-2501-AWQ" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/stelterlab/Mistral-Small-24B-Instruct-2501-AWQ/raw/cbda099649a0188dd888d44f0e4964d8d982dc9a/model.safetensors.index.json" } ] }, { "id": "MaziyarPanahi/Qwen3-14B-GGUF", "name": "Qwen3-14B-GGUF GGUF F16", "short_name": "Qwen3 14B", "license": "unknown", "hf_pipeline_tag": "text-generation", "hf_downloads": 229212, "tags": [ "gguf", "mistral", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:Qwen/Qwen3-14B", "base_model:quantized:Qwen/Qwen3-14B", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 GGUF (Qwen3ForCausalLM)", "total_params_b": 14.7683072, "active_params_b": 13.99039488, "layers": 40, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF F16", "weight_bytes_per_param": 2.000461078166088, "resident_weight_gb": 29.543423744, "active_weight_gb": 27.98163968, "kv_alloc_gb_per_1k_tokens": 0.16384, "kv_read_gb_per_1k_tokens": 0.16384, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, model card, audited BF16 base profile comparison, linked GGUF file sizes, stale repo config check, and direct selected F16 GGUF tensor-index range read. The authoritative bounds profile uses Qwen3-14B.fp16.gguf because it matches API gguf.totalFileSize: 29.543423744 GB resident and 27.981639680 GB ordinary text-decode swept traffic after excluding token_embd.weight." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-14B-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-14B-GGUF/raw/421b77e6e553db95b5ca5eae8055b2677bc4f41f/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-14B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Qwen3-14B-GGUF" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-14B-GGUF/resolve/421b77e6e553db95b5ca5eae8055b2677bc4f41f/Qwen3-14B.fp16.gguf" } ] }, { "id": "nvidia/Qwen3-32B-NVFP4", "name": "Qwen3-32B-NVFP4 NVFP4", "short_name": "Qwen3 32B NVFP4", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 246588, "tags": [ "Model Optimizer", "safetensors", "qwen3", "nvidia", "ModelOpt", "Qwen3", "quantized", "FP4", "fp4", "text-generation", "conversational", "base_model:Qwen/Qwen3-32B", "base_model:quantized:Qwen/Qwen3-32B", "license:apache-2.0", "8-bit", "modelopt", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 32.762123264, "active_params_b": 31.984210944, "layers": 64, "attention_heads": 64, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "ModelOpt NVFP4 + FP8 scales + BF16 fixed tensors", "weight_bytes_per_param": 0.5, "resident_weight_gb": 20.666169344, "active_weight_gb": 19.110344704, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, NVIDIA model card, pinned served config, ModelOpt quantization config, current Qwen3 32B base config comparison, safetensors index, linked-object HEAD checks, and direct safetensors shard-header range reads. Legacy fallback values mirror the audited profile: exact 20.666169344 GB stored resident package, 19.110344704 GB ordinary text-decode swept traffic, 1.555824640 GB resident-only input embedding, and FP8 full-context KV. The authoritative profile supersedes the generated simple 0.5-byte NVFP4 resident estimate and BF16 KV estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/Qwen3-32B-NVFP4" }, { "label": "Model config", "url": "https://huggingface.co/nvidia/Qwen3-32B-NVFP4/raw/16426c6eb87be9e27c14cc9fb318f9c7a5f8588c/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-32B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/Qwen3-32B-NVFP4" }, { "label": "ModelOpt quantization config", "url": "https://huggingface.co/nvidia/Qwen3-32B-NVFP4/raw/16426c6eb87be9e27c14cc9fb318f9c7a5f8588c/hf_quant_config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/nvidia/Qwen3-32B-NVFP4/raw/16426c6eb87be9e27c14cc9fb318f9c7a5f8588c/model.safetensors.index.json" } ] }, { "id": "MaziyarPanahi/Qwen3-8B-GGUF", "name": "Qwen3-8B-GGUF GGUF 2-bit", "short_name": "Qwen3 8B", "license": "unknown", "hf_pipeline_tag": "text-generation", "hf_downloads": 227628, "tags": [ "gguf", "mistral", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:Qwen/Qwen3-8B", "base_model:quantized:Qwen/Qwen3-8B", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 GGUF selected F16 artifact", "total_params_b": 8.19073536, "active_params_b": 7.568405504, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF F16 selected artifact", "weight_bytes_per_param": 2.000802526233737, "resident_weight_gb": 16.388044, "active_weight_gb": 15.137427456, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, pinned model card, stale repo-config check, audited BF16 Qwen3 8B base profile, linked-object HEAD checks, and direct GGUF header range read. Legacy fallback values mirror the audited profile: API-selected Qwen3-8B.fp16.gguf, exact 16.388044000 GB resident linked file, 15.137427456 GB ordinary text-decode swept tensor traffic, 1.250616544 GB resident-only input embedding plus GGUF overhead, and FP16 full-context KV. The authoritative profile supersedes the generated mistral/2048-context/2-bit estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-8B-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-8B-GGUF/raw/ac6dd95cf227fe9138362c0536fe3c3802008ccf/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-8B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Qwen3-8B-GGUF" }, { "label": "Selected F16 GGUF artifact", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-8B-GGUF/resolve/ac6dd95cf227fe9138362c0536fe3c3802008ccf/Qwen3-8B.fp16.gguf" } ] }, { "id": "MaziyarPanahi/Qwen3-4B-GGUF", "name": "Qwen3-4B-GGUF GGUF 2-bit", "short_name": "Qwen3 4B", "license": "unknown", "hf_pipeline_tag": "text-generation", "hf_downloads": 226707, "tags": [ "gguf", "mistral", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:Qwen/Qwen3-4B", "base_model:quantized:Qwen/Qwen3-4B", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 GGUF selected F16 artifact", "total_params_b": 4.022468096, "active_params_b": 4.022468096, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF F16 selected artifact", "weight_bytes_per_param": 2.0015783742340463, "resident_weight_gb": 8.051285152, "active_weight_gb": 8.045328384, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, pinned model card, stale repo-config check, audited BF16 Qwen3 4B base profile, linked-object HEAD checks, and direct GGUF header range read. Legacy fallback values mirror the audited profile: API-selected Qwen3-4B.fp16.gguf, exact 8.051285152 GB resident linked file, 8.045328384 GB ordinary text-decode swept tensor traffic, 0.005956768 GB resident-only GGUF overhead, tied token_embd output projection, and FP16 full-context KV. The authoritative profile supersedes the generated mistral/2048-context/2-bit estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-4B-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-4B-GGUF/raw/b507d8f2622a6be905066595a5b3cd00e11ab379/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-4B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Qwen3-4B-GGUF" }, { "label": "Selected F16 GGUF artifact", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-4B-GGUF/resolve/b507d8f2622a6be905066595a5b3cd00e11ab379/Qwen3-4B.fp16.gguf" } ] }, { "id": "MaziyarPanahi/Qwen3-32B-GGUF", "name": "Qwen3-32B-GGUF GGUF 2-bit", "short_name": "Qwen3 32B", "license": "unknown", "hf_pipeline_tag": "text-generation", "hf_downloads": 225531, "tags": [ "gguf", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:Qwen/Qwen3-32B", "base_model:quantized:Qwen/Qwen3-32B", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 GGUF selected Q2_K artifact", "total_params_b": 32.762123264, "active_params_b": 31.984210944, "layers": 64, "attention_heads": 64, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q2_K selected artifact", "weight_bytes_per_param": 0.37679644522809885, "resident_weight_gb": 12.344651584, "active_weight_gb": 12.083424256, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, pinned model card, missing repo-config check, audited BF16 Qwen3 32B base profile, linked-object HEAD checks, and direct GGUF header range read. Legacy fallback values mirror the audited profile: API-selected Qwen3-32B.Q2_K.gguf, exact 12.344651584 GB resident linked file, 12.083424256 GB ordinary text-decode swept tensor traffic, 0.261227328 GB resident-only input embedding plus GGUF overhead, and FP16 full-context KV. The selected artifact mixes Q2_K, Q3_K, Q4_K, Q6_K, and F32 tensor classes, so the authoritative profile supersedes the generated flat 2-bit estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-32B-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3-32B/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-32B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Qwen3-32B-GGUF" }, { "label": "Selected Q2_K GGUF artifact", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-32B-GGUF/resolve/0dd9f2cba883d7114816444ae9ee9219cf87a917/Qwen3-32B.Q2_K.gguf" } ] }, { "id": "lmstudio-community/gemma-4-26B-A4B-it-MLX-4bit", "name": "gemma-4-26B-A4B-it-MLX-4bit 4-bit", "short_name": "gemma 4 26B A4B it MLX 4bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 234475, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "mlx", "conversational", "base_model:google/gemma-4-26B-A4B-it", "base_model:quantized:google/gemma-4-26B-A4B-it", "license:apache-2.0", "endpoints_compatible", "4-bit", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "gemma4_text", "total_params_b": 26.544131376, "active_params_b": 4, "layers": 30, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "MLX 4-bit affine", "weight_bytes_per_param": 0.5880250448923184, "resident_weight_gb": 15.608614044, "active_weight_gb": 2.424219708, "kv_alloc_gb_per_1k_tokens": 0.20992, "kv_read_gb_per_1k_tokens": 0.20992, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, served MLX config, non-QAT Google config comparison, safetensors index, and direct range-read shard headers. Bound weights separate the exact 15.608614044 GB resident MLX package from 1.621321788 GB fixed ordinary text traffic plus 0.100362240 GB per routed expert index; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-MLX-4bit" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-26B-A4B-it-MLX-4bit" }, { "label": "Pinned MLX model config", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-MLX-4bit/raw/3af5252ed1e675e6bba9be8cc3087bc00920799c/config.json" }, { "label": "Pinned Google base config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it/raw/20da991ab4afab98e8f910c4a2e8f4fbefc404ad/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-MLX-4bit/raw/3af5252ed1e675e6bba9be8cc3087bc00920799c/model.safetensors.index.json" } ] }, { "id": "MaziyarPanahi/Qwen3-1.7B-GGUF", "name": "Qwen3-1.7B-GGUF GGUF 2-bit", "short_name": "Qwen3 1.7B", "license": "unknown", "hf_pipeline_tag": "text-generation", "hf_downloads": 224978, "tags": [ "gguf", "mistral", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:Qwen/Qwen3-1.7B", "base_model:quantized:Qwen/Qwen3-1.7B", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 GGUF selected F16 artifact", "total_params_b": 2.031739904, "active_params_b": 1.720574976, "layers": 28, "attention_heads": 16, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF F16 selected artifact", "weight_bytes_per_param": 2.0030512882026854, "resident_weight_gb": 4.069679232, "active_weight_gb": 3.44139776, "kv_alloc_gb_per_1k_tokens": 0.114688, "kv_read_gb_per_1k_tokens": 0.114688, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, pinned model card, stale repo-config check, audited BF16 Qwen3 1.7B parent profile, linked-object HEAD checks, and direct GGUF header range read. Legacy fallback values mirror the audited profile: API-selected Qwen3-1.7B.fp16.gguf, exact 4.069679232 GB resident linked file, 3.441397760 GB ordinary text-decode swept tensor traffic, 0.628281472 GB resident-only input embedding plus GGUF overhead, and FP16 full-context KV. The authoritative profile supersedes the generated mistral/2048-context/2-bit estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-1.7B-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-1.7B-GGUF/raw/57a69d3bba4c7254d3466a3bb96d909b988f8fbd/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-1.7B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Qwen3-1.7B-GGUF" }, { "label": "Selected F16 GGUF artifact", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-1.7B-GGUF/resolve/57a69d3bba4c7254d3466a3bb96d909b988f8fbd/Qwen3-1.7B.fp16.gguf" } ] }, { "id": "MaziyarPanahi/Qwen3-30B-A3B-GGUF", "name": "Qwen3-30B-A3B-GGUF GGUF 2-bit", "short_name": "Qwen3 30B A3B", "license": "unknown", "hf_pipeline_tag": "text-generation", "hf_downloads": 222592, "tags": [ "gguf", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:Qwen/Qwen3-30B-A3B", "base_model:quantized:Qwen/Qwen3-30B-A3B", "region:us", "conversational" ], "architecture": { "type": "MoE transformer", "detail": "qwen3moe GGUF selected Q2_K artifact", "total_params_b": 30.532122624, "active_params_b": 3.041867776, "layers": 48, "attention_heads": 32, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "GGUF Q2_K selected artifact", "weight_bytes_per_param": 0.3687463770091794, "resident_weight_gb": 11.2586096, "active_weight_gb": 1.312274432, "expert_param_b": 0.226492416, "fixed_param_b": 1.229928448, "shared_expert_param_b": 0, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "notes": "Manually audited on 2026-07-06 from current HF API metadata, linked-object sizes, audited BF16 parent profile, and a selected Q2_K GGUF tensor-index range read. The HF API-selected artifact is Qwen3-30B-A3B.Q2_K.gguf at 11.258609600 GB. Ordinary text decode charges exact fixed traffic plus 8 expected-distinct routed expert byte groups: 0.656390144 GB fixed + 8 * 0.081985536 GB routed = 1.312274432 GB swept traffic. KV uses full-context FP16 llama.cpp-style cache from the GGUF/base geometry." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-30B-A3B-GGUF/raw/6443554b0ef1176268cae69212105bf4c7187091/README.md" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Qwen3-30B-A3B-GGUF" }, { "label": "Audited BF16 parent config", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B/raw/ad44e777bcd18fa416d9da3bd8f70d33ebb85d39/config.json" }, { "label": "Selected Q2_K GGUF artifact", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-30B-A3B-GGUF/resolve/6443554b0ef1176268cae69212105bf4c7187091/Qwen3-30B-A3B.Q2_K.gguf" } ] }, { "id": "Qwen/Qwen3.5-0.8B-Base", "name": "Qwen3.5-0.8B-Base F32/BF16", "short_name": "Qwen3.5 0.8B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 283029, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "conversational", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text hybrid full-attention/DeltaNet", "total_params_b": 0.873438784, "active_params_b": 0.752393024, "layers": 24, "attention_heads": 8, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "F32/BF16", "weight_bytes_per_param": 2.000005935161221, "resident_weight_gb": 1.746882752, "active_weight_gb": 1.504791232, "kv_alloc_gb_per_1k_tokens": 0.012288, "kv_read_gb_per_1k_tokens": 0.012288, "kv_alloc_fixed_gb": 0.019759104, "kv_read_fixed_gb": 0.019759104, "notes": "Manually audited on 2026-07-06 from HF API commit dc7cdfe2ee4154fa7e30f5b51ca41bfa40174e68, model card, config.json, safetensors index, direct safetensors header range read, and the Transformers Qwen3.5 runtime implementation. Ordinary text decode sweeps exact model.language_model tensor bytes: 1.504791232 GB. Visual and MTP tensors remain resident-only. Text KV uses six full-attention layers at 0.012288 GB per 1K tokens plus a fixed DeltaNet recurrent/conv state charge of 0.019759104 GB per session/token." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3.5-0.8B-Base/raw/dc7cdfe2ee4154fa7e30f5b51ca41bfa40174e68/README.md" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3.5-0.8B-Base/raw/dc7cdfe2ee4154fa7e30f5b51ca41bfa40174e68/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-0.8B-Base" }, { "label": "Safetensors index", "url": "https://huggingface.co/Qwen/Qwen3.5-0.8B-Base/raw/dc7cdfe2ee4154fa7e30f5b51ca41bfa40174e68/model.safetensors.index.json" } ] }, { "id": "allenai/wildguard", "name": "wildguard BF16", "short_name": "wildguard", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 264728, "tags": [ "transformers", "pytorch", "safetensors", "mistral", "text-generation", "classifier", "safety", "moderation", "llm", "lm", "en", "dataset:allenai/wildguardmix", "arxiv:2406.18495", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "mistral gated config unavailable", "total_params_b": 7.248031744, "active_params_b": 7.248031744, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 14.496063488, "active_weight_gb": 14.496063488, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manually reviewed on 2026-07-06 from HF API metadata, authenticated model-card access, raw access checks, and Mistral 7B v0.3 lineage context. The authoritative model profile intentionally fails closed: WildGuard config.json and model.safetensors.index.json are gated and inaccessible to the configured HF CLI identity, so KV geometry, exact context, tied embeddings, vocabulary changes, and swept text-decode traffic are not verified." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/allenai/wildguard" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/allenai/wildguard" }, { "label": "Gated config access check", "url": "https://huggingface.co/allenai/wildguard/raw/cbba4823f3e8020e5a74a5e29bf85072def6f2ff/config.json" } ] }, { "id": "lmstudio-community/gemma-4-26B-A4B-it-QAT-GGUF", "name": "gemma-4-26B-A4B-it-QAT-GGUF GGUF 4-bit", "short_name": "gemma 4 26B A4B it QAT", "license": "apache-2.0", "hf_pipeline_tag": "unknown", "hf_downloads": 302235, "tags": [ "gguf", "base_model:google/gemma-4-26B-A4B-it-qat-q4_0-unquantized", "base_model:quantized:google/gemma-4-26B-A4B-it-qat-q4_0-unquantized", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "MoE transformer", "detail": "gemma4_text", "total_params_b": 25.233142046, "active_params_b": 3.8, "layers": 30, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "GGUF Q4_0 QAT", "weight_bytes_per_param": 0.5722380005501119, "resident_weight_gb": 14.439362752, "active_weight_gb": 2.380070008, "kv_alloc_gb_per_1k_tokens": 0.20992, "kv_read_gb_per_1k_tokens": 0.20992, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, pinned model card, selected GGUF linked-object HEAD checks, Google QAT base config, and direct range-read GGUF header/tensor index. Bound weights separate the exact 14.439362752 GB resident selected file from 1.577172088 GB fixed ordinary text traffic plus 0.100362240 GB per routed expert index; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-QAT-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-26B-A4B-it-QAT-GGUF" }, { "label": "Public Google QAT reference config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it-qat-q4_0-unquantized/raw/641f184470aa8554ae7957599a624badc2bf4e57/config.json" }, { "label": "Selected GGUF header audit", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-QAT-GGUF/resolve/43abd4da26c7b9c1a93c4d96a588c8a5374be8eb/gemma-4-26B-A4B-it-QAT-Q4_0.gguf" } ] }, { "id": "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF", "name": "Meta-Llama-3.1-8B-Instruct-GGUF GGUF", "short_name": "Meta Llama 3.1 8B Instruct", "license": "llama3.1", "hf_pipeline_tag": "text-generation", "hf_downloads": 261959, "tags": [ "gguf", "facebook", "meta", "pytorch", "llama", "llama-3", "text-generation", "en", "de", "fr", "it", "pt", "hi", "es", "th", "base_model:meta-llama/Llama-3.1-8B-Instruct", "base_model:quantized:meta-llama/Llama-3.1-8B-Instruct", "license:llama3.1", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "llama GGUF selected F32 artifact", "total_params_b": 8.030261312, "active_params_b": 7.504924736, "layers": 32, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF F32 selected artifact", "weight_bytes_per_param": 4.000976386657341, "resident_weight_gb": 32.128885888, "active_weight_gb": 30.019698944, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, pinned model card, gated base-profile checks, linked-object HEAD checks, missing repo-config check, and direct range-read GGUF header/tensor index. The API-selected artifact is Meta-Llama-3.1-8B-Instruct-f32.gguf: exact 32.128885888 GB resident linked file, 30.019698944 GB ordinary text-decode swept tensor traffic, 2.109186944 GB resident-only input embedding plus GGUF overhead, and FP16 full-context KV. The audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF" }, { "label": "Selected GGUF header audit", "url": "https://huggingface.co/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF/resolve/bf5b95e96dac0462e2a09145ec66cae9a3f12067/Meta-Llama-3.1-8B-Instruct-f32.gguf" }, { "label": "Gated base model card", "url": "https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct" } ] }, { "id": "zai-org/GLM-4.6V-Flash", "name": "GLM-4.6V-Flash BF16", "short_name": "GLM 4.6V Flash", "license": "mit", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 343891, "tags": [ "transformers", "safetensors", "glm4v", "image-text-to-text", "conversational", "zh", "en", "arxiv:2507.01006", "license:mit", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "glm4v_text", "total_params_b": 10.292777472, "active_params_b": 8.779522048, "layers": 40, "attention_heads": 32, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 20.585554944, "active_weight_gb": 17.559044096, "kv_alloc_gb_per_1k_tokens": 0.04096, "kv_read_gb_per_1k_tokens": 0.04096, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, model card, pinned served GLM4V config, Transformers GLM4V implementation review, safetensors index, and direct shard header range reads. Bound weights separate the exact 20.585554944 GB resident BF16 multimodal package from 17.559044096 GB ordinary text-decode swept traffic; model.visual.* and model.language_model.embed_tokens.weight are resident-only for ordinary text-token decode. The audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/zai-org/GLM-4.6V-Flash" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/zai-org/GLM-4.6V-Flash" }, { "label": "Pinned model config", "url": "https://huggingface.co/zai-org/GLM-4.6V-Flash/raw/411bb4d77144a3f03accbf4b780f5acb8b7cde4e/config.json" }, { "label": "Transformers GLM4V implementation", "url": "https://raw.githubusercontent.com/huggingface/transformers/v5.0.0rc0/src/transformers/models/glm4v/modeling_glm4v.py" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/zai-org/GLM-4.6V-Flash/resolve/411bb4d77144a3f03accbf4b780f5acb8b7cde4e/model.safetensors.index.json" } ] }, { "id": "yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-GGUF", "name": "gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-GGUF GGUF", "short_name": "gemma 4 12B agentic fable5 composer2.5 v2 3.5x t", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 355871, "tags": [ "gguf", "gemma4", "coding", "agentic", "terminal", "tool-use", "reasoning", "thinking", "llama.cpp", "local-llm", "text-generation", "base_model:google/gemma-4-12B-it", "base_model:quantized:google/gemma-4-12B-it", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_unified_text", "total_params_b": 11.907350576, "active_params_b": 11.907350576, "layers": 48, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q4_K_M", "weight_bytes_per_param": 0.6199012632480672, "resident_weight_gb": 7.381381664, "active_weight_gb": 7.365559808, "kv_alloc_gb_per_1k_tokens": 0.335872, "kv_read_gb_per_1k_tokens": 0.335872, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, pinned model card, missing repo-local config check, Google base config, linked GGUF file sizes, and a direct GGUF tensor-index range read. The authoritative bounds profile uses the card-recommended gemma4-v2-Q4_K_M.gguf artifact: 7.381381664 GB resident, 7.365559808 GB swept text traffic, 0.015821856 GB resident-only GGUF metadata/header bytes, and layered Gemma 4 hybrid KV. API gguf.totalFileSize selects the smaller Q3_K_M file, but the model card calls Q4_K_M the recommended sweet spot and uses it in the example command. Legacy per-1k KV fields show the short-context unsaturated coefficient; long-context calculation must use the profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-GGUF" }, { "label": "Google base config", "url": "https://huggingface.co/google/gemma-4-12B-it/raw/5926caa4ec0cac5cbfadaf4077420520de1d5205/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-GGUF" }, { "label": "Selected GGUF artifact", "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" } ] }, { "id": "Qwen/Qwen2.5-Math-1.5B", "name": "Qwen2.5-Math-1.5B BF16", "short_name": "Qwen2.5 Math 1.5B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 207753, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "conversational", "en", "arxiv:2409.12122", "base_model:Qwen/Qwen2.5-1.5B", "base_model:finetune:Qwen/Qwen2.5-1.5B", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 1.543714304, "active_params_b": 1.543714304, "layers": 28, "attention_heads": 12, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 4096, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 3.087428608, "active_weight_gb": 3.087428608, "kv_alloc_gb_per_1k_tokens": 0.028672, "kv_read_gb_per_1k_tokens": 0.028672, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, pinned served config, Qwen2.5 1.5B base config comparison, and a direct safetensors header range read. Exact BF16 tensor payload is 3.087428608 GB / 1.543714304B parameters; tie_word_embeddings is true and no lm_head.weight is stored, so the full tensor payload is charged for ordinary decode." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Math-1.5B" }, { "label": "Pinned model config", "url": "https://huggingface.co/Qwen/Qwen2.5-Math-1.5B/raw/4a83ca6e4526a4f2da3aa259ec36c259f66b2ab2/config.json" }, { "label": "Qwen2.5 1.5B base config", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B/raw/8faed761d45a263340a0528343f099c05c9a4323/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Math-1.5B" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/Qwen/Qwen2.5-Math-1.5B/resolve/4a83ca6e4526a4f2da3aa259ec36c259f66b2ab2/model.safetensors" } ] }, { "id": "Jackrong/Qwopus3.6-27B-v2-MTP-GGUF", "name": "Qwopus3.6-27B-v2-MTP-GGUF GGUF", "short_name": "Qwopus3.6 27B v2 MTP", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 168444, "tags": [ "transformers", "gguf", "llama.cpp", "image-text-to-text", "vision", "multimodal", "text-generation-inference", "unsloth", "conversational", "mtp", "multi-token-prediction", "speculative-decoding", "qwen3_6", "reasoning", "chain-of-thought", "lora", "sft", "agent", "coder", "devops", "math", "science", "image", "tool-use", "function-calling", "long-context", "en", "zh", "ko", "ru", "ja", "es", "dataset:Jackrong/Claude-opus-4.7-TraceInversion-5000x", "dataset:Jackrong/Claude-opus-4.6-TraceInversion-9000x", "base_model:Qwen/Qwen3.6-27B", "base_model:adapter:Qwen/Qwen3.6-27B", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 27.320697856, "active_params_b": 25.624600064, "layers": 64, "attention_heads": 24, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q5_K_M", "weight_bytes_per_param": 0.7150512493848942, "resident_weight_gb": 19.535699136, "active_weight_gb": 18.346018816, "kv_alloc_gb_per_1k_tokens": 0.065536, "kv_read_gb_per_1k_tokens": 0.065536, "kv_alloc_fixed_gb": 0.154927104, "kv_read_fixed_gb": 0.154927104, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, model card, Qwen3.6 base config comparison, selected linked-object HEAD checks, direct GGUF header/tensor-index range read, and Qwen3.5 runtime review. The API GGUF block points at mmproj-F32, so the audited profile uses the selected Q5_K_M main artifact and supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Jackrong/Qwopus3.6-27B-v2-MTP-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Jackrong/Qwopus3.6-27B-v2-MTP-GGUF" }, { "label": "Selected Q5_K_M GGUF artifact", "url": "https://huggingface.co/Jackrong/Qwopus3.6-27B-v2-MTP-GGUF/resolve/de55021a5b03540dc4a5ceec6bffff00e62fde65/Qwopus3.6-27B-v2-MTP-Q5_K_M.gguf" }, { "label": "Qwen3.6 base config", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json" }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py" } ] }, { "id": "Qwen/Qwen3-Next-80B-A3B-Instruct", "name": "Qwen3-Next-80B-A3B-Instruct BF16", "short_name": "Qwen3 Next 80B A3B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 243964, "tags": [ "transformers", "safetensors", "qwen3_next", "text-generation", "conversational", "arxiv:2309.00071", "arxiv:2404.06654", "arxiv:2505.09388", "arxiv:2501.15383", "license:apache-2.0", "eval-results", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_next (Qwen3NextForCausalLM)", "total_params_b": 81.32486272, "active_params_b": 3.56376448, "layers": 48, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 162.64972544, "active_weight_gb": 7.12752896, "kv_alloc_gb_per_1k_tokens": 0.024576, "kv_read_gb_per_1k_tokens": 0.024576, "kv_alloc_fixed_gb": 0.077856768, "kv_read_fixed_gb": 0.077856768, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, model card, served config, safetensors index, direct safetensors header range reads, and Qwen3-Next runtime review. Active traffic is fixed decode traffic plus 10 routed expert groups; top-level MTP tensors are resident-only for non-speculative decode. The audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct/raw/9c7f2fbe84465e40164a94cc16cd30b6999b0cc7/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-Next-80B-A3B-Instruct" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct/resolve/9c7f2fbe84465e40164a94cc16cd30b6999b0cc7/model.safetensors.index.json" }, { "label": "Transformers Qwen3-Next implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/qwen3_next/modeling_qwen3_next.py" } ] }, { "id": "QuantTrio/Qwen3.5-27B-AWQ", "name": "Qwen3.5-27B-AWQ 4-bit", "short_name": "Qwen3.5 27B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 213771, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "vLLM", "AWQ", "conversational", "base_model:Qwen/Qwen3.5-27B", "base_model:quantized:Qwen/Qwen3.5-27B", "license:apache-2.0", "endpoints_compatible", "4-bit", "awq", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 27.7814, "active_params_b": 27.7814, "layers": 64, "attention_heads": 24, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "4-bit", "weight_bytes_per_param": 0.7865993885467936, "resident_weight_gb": 21.85285424, "active_weight_gb": 17.539198464, "kv_alloc_gb_per_1k_tokens": 0.065536, "kv_read_gb_per_1k_tokens": 0.065536, "kv_alloc_fixed_gb": 0.154927104, "kv_read_fixed_gb": 0.154927104, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, model card, served AWQ config, safetensors index, direct safetensors header range reads, base config comparison, and Qwen3.5 runtime review. Resident weights are exact stored tensor payloads; active traffic is ordinary text decode through language layers/norm and lm_head; resident-only input embedding, visual, and MTP tensors are excluded from per-token swept traffic. The audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/QuantTrio/Qwen3.5-27B-AWQ" }, { "label": "Model config", "url": "https://huggingface.co/QuantTrio/Qwen3.5-27B-AWQ/raw/56f41874389615226dcd849ded92261a0286ff59/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-27B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/QuantTrio/Qwen3.5-27B-AWQ" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/QuantTrio/Qwen3.5-27B-AWQ/resolve/56f41874389615226dcd849ded92261a0286ff59/model.safetensors.index.json" }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py" } ] }, { "id": "deepseek-ai/DeepSeek-V3.1", "name": "DeepSeek-V3.1 FP8", "short_name": "DeepSeek V3.1", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 250886, "tags": [ "transformers", "safetensors", "deepseek_v3", "text-generation", "conversational", "custom_code", "arxiv:2412.19437", "base_model:deepseek-ai/DeepSeek-V3.1-Base", "base_model:quantized:deepseek-ai/DeepSeek-V3.1-Base", "license:mit", "eval-results", "text-generation-inference", "endpoints_compatible", "fp8", "region:us" ], "architecture": { "type": "MLA MoE transformer", "detail": "deepseek_v3 (DeepseekV3ForCausalLM)", "total_params_b": 671, "active_params_b": 37, "layers": 61, "attention_heads": 128, "kv_heads": 128, "head_dim": 128, "max_context_tokens": 163840, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "kv_lora_rank": 512, "qk_rope_head_dim": 64, "qk_nope_head_dim": 128, "v_head_dim": 128 }, "adapter": { "kind": "moe", "weight_precision": "FP8", "weight_bytes_per_param": 1, "resident_weight_gb": 688.57483936, "active_weight_gb": 39.521833312, "kv_alloc_gb_per_1k_tokens": 0.070272, "kv_read_gb_per_1k_tokens": 0.070272, "expert_param_b": 2.554954752, "shared_expert_param_b": 2.554954752, "fixed_param_b": 16.527240544, "notes": "Manual correction on 2026-07-06 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 headers. Bound weights use exact stored FP8/BF16/F32 shard bytes and official absorb-MLA cache geometry; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.1" }, { "label": "Model config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.1/raw/c0781d039fb7a1ba2abc4add0bdc293e92d2b8db/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.1-Base" }, { "label": "Base model config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.1-Base/raw/d3d4eafdc470de44bbf6f0a74f852eb522357be8/config.json" }, { "label": "DeepSeek-V3 local inference code", "url": "https://github.com/deepseek-ai/DeepSeek-V3/blob/9b4e9788e4a3a731f7567338ed15d3ec549ce03b/inference/model.py" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-V3.1" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.1/resolve/c0781d039fb7a1ba2abc4add0bdc293e92d2b8db/model.safetensors.index.json" } ] }, { "id": "leon-se/gemma-4-E4B-it-FP8-Dynamic", "name": "gemma-4-E4B-it-FP8-Dynamic FP8", "short_name": "gemma 4 E4B it FP8 Dynamic", "license": "unknown", "hf_pipeline_tag": "unknown", "hf_downloads": 182449, "tags": [ "safetensors", "gemma4", "base_model:google/gemma-4-E4B-it", "base_model:quantized:google/gemma-4-E4B-it", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 8.66860081, "active_params_b": 4.700851498, "layers": 42, "attention_heads": 8, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "FP8 dynamic", "weight_bytes_per_param": 1, "resident_weight_gb": 13.30935906, "active_weight_gb": 5.373860436, "kv_alloc_gb_per_1k_tokens": 0.028672, "kv_read_gb_per_1k_tokens": 0.028672, "kv_alloc_fixed_gb": 0.03670016, "kv_read_fixed_gb": 0.03670016, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, served compressed-tensors config, recipe.yaml, base Google config comparison, linked-object HEAD check, and direct safetensors header range read. Bound weights separate the exact 13.309359060 GB resident FP8/BF16 package from the 5.373860436 GB ordinary text-decode swept subset; hybrid Gemma 4 E4B KV uses full-context plus fixed sliding-window cache terms. The audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/leon-se/gemma-4-E4B-it-FP8-Dynamic" }, { "label": "Model config", "url": "https://huggingface.co/leon-se/gemma-4-E4B-it-FP8-Dynamic/raw/56e30bf603d18a4972caffafa1bb4a4f9a841dee/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-E4B-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/leon-se/gemma-4-E4B-it-FP8-Dynamic" }, { "label": "Quantization recipe", "url": "https://huggingface.co/leon-se/gemma-4-E4B-it-FP8-Dynamic/raw/56e30bf603d18a4972caffafa1bb4a4f9a841dee/recipe.yaml" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/leon-se/gemma-4-E4B-it-FP8-Dynamic/resolve/56e30bf603d18a4972caffafa1bb4a4f9a841dee/model.safetensors" } ] }, { "id": "nvidia/NVIDIA-Nemotron-Nano-9B-v2-FP8", "name": "NVIDIA-Nemotron-Nano-9B-v2-FP8 FP8", "short_name": "NVIDIA Nemotron Nano 9B v2 FP8", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 227137, "tags": [ "transformers", "safetensors", "nemotron_h", "text-generation", "nvidia", "pytorch", "conversational", "custom_code", "en", "es", "fr", "de", "it", "ja", "dataset:nvidia/Nemotron-Post-Training-Dataset-v1", "dataset:nvidia/Nemotron-Post-Training-Dataset-v2", "dataset:nvidia/Nemotron-Pretraining-Dataset-sample", "dataset:nvidia/Nemotron-CC-v2", "dataset:nvidia/Nemotron-CC-Math-v1", "dataset:nvidia/Nemotron-Pretraining-SFT-v1", "arxiv:2504.03624", "arxiv:2508.14444", "arxiv:2412.02595", "base_model:nvidia/NVIDIA-Nemotron-Nano-9B-v2", "base_model:finetune:nvidia/NVIDIA-Nemotron-Nano-9B-v2", "license:other", "endpoints_compatible", "region:us" ], "architecture": { "type": "Hybrid Mamba/Transformer", "detail": "nemotron_h (NemotronHForCausalLM); 27 Mamba, 4 attention, 25 MLP layers", "total_params_b": 8.888227536, "active_params_b": 8.301024976, "layers": 56, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "hybrid", "weight_precision": "FP8/BF16/F32", "weight_bytes_per_param": 1, "resident_weight_gb": 10.285035328, "active_weight_gb": 9.110630208, "kv_alloc_gb_per_1k_tokens": 0.016384, "kv_read_gb_per_1k_tokens": 0.016384, "kv_alloc_fixed_gb": 0.144211968, "kv_read_fixed_gb": 0.144211968, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, model card, served config, audited BF16 config comparison, ModelOpt quantization config, custom runtime code, safetensors index, and direct range-read safetensors shard headers. Bound weights separate the exact 10.285035328 GB resident ModelOpt package from the 9.110630208 GB ordinary text-decode swept subset. hf_quant_config.json records kv_cache_quant_algo null and the vLLM card command requires --mamba_ssm_cache_dtype float32, so legacy fallback values use BF16 attention KV plus fixed BF16-convolution/F32-SSM state. The audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2-FP8" }, { "label": "Model config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2-FP8/raw/8bc5eece2eb5514c4bca7f2ec655b91eb554f4c0/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/NVIDIA-Nemotron-Nano-9B-v2-FP8" }, { "label": "ModelOpt quantization config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2-FP8/raw/8bc5eece2eb5514c4bca7f2ec655b91eb554f4c0/hf_quant_config.json" }, { "label": "Custom runtime code", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2-FP8/raw/8bc5eece2eb5514c4bca7f2ec655b91eb554f4c0/modeling_nemotron_h.py" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2-FP8/resolve/8bc5eece2eb5514c4bca7f2ec655b91eb554f4c0/model.safetensors.index.json" } ] }, { "id": "lmstudio-community/GLM-4.7-Flash-MLX-8bit", "name": "GLM-4.7-Flash-MLX-8bit 8-bit", "short_name": "GLM 4.7 Flash MLX 8bit", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 272241, "tags": [ "transformers", "safetensors", "glm4_moe_lite", "text-generation", "mlx", "conversational", "en", "zh", "base_model:zai-org/GLM-4.7-Flash", "base_model:quantized:zai-org/GLM-4.7-Flash", "license:mit", "endpoints_compatible", "8-bit", "region:us" ], "architecture": { "type": "Latent MLA MoE transformer", "detail": "glm4_moe_lite (Glm4MoeLiteForCausalLM)", "total_params_b": 29.94339392, "active_params_b": 3.584459991, "layers": 47, "attention_heads": 20, "kv_heads": 20, "head_dim": 256, "max_context_tokens": 202752, "routed_experts": 64, "routed_experts_per_token": 4, "shared_experts_per_token": 1, "kv_lora_rank": 512, "qk_rope_head_dim": 64, "qk_nope_head_dim": 192, "v_head_dim": 256 }, "adapter": { "kind": "moe", "weight_precision": "MLX 8-bit affine", "weight_bytes_per_param": 1.0626970360479431, "resident_weight_gb": 31.820755968, "active_weight_gb": 3.809195008, "kv_alloc_gb_per_1k_tokens": 0.054144, "kv_read_gb_per_1k_tokens": 0.054144, "expert_param_b": 0.4340299750108565, "shared_expert_param_b": 0.4340299750108565, "fixed_param_b": 1.4143101156934004, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, pinned served MLX config, pinned BF16 GLM source config comparison, safetensors index, direct shard header reads, and MLX-LM glm4_moe_lite runtime review. Header bytes record 31.820755968 GB resident, 31.483737088 GB ordinary text resident, 29.519511552 GB packed routed expert tensors, 1.964225536 GB fixed ordinary text traffic, and 0.33701888 GB resident-only embedding tensors. MLX-LM stores latent MLA cache state at 0.054144 GB per 1k context tokens; the audited profile supersedes this legacy adapter summary for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/GLM-4.7-Flash-MLX-8bit" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/GLM-4.7-Flash-MLX-8bit" }, { "label": "Pinned model config", "url": "https://huggingface.co/lmstudio-community/GLM-4.7-Flash-MLX-8bit/raw/0eca07a0c2944693945f21a18cf663722e3186b9/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/zai-org/GLM-4.7-Flash" }, { "label": "Pinned base config", "url": "https://huggingface.co/zai-org/GLM-4.7-Flash/raw/7dd20894a642a0aa287e9827cb1a1f7f91386b67/config.json" }, { "label": "Safetensors shard header audit", "url": "https://huggingface.co/lmstudio-community/GLM-4.7-Flash-MLX-8bit/raw/0eca07a0c2944693945f21a18cf663722e3186b9/model.safetensors.index.json" }, { "label": "MLX-LM GLM4 MoE Lite implementation", "url": "https://github.com/ml-explore/mlx-lm/blob/2ed22318cd6a2fcc5c2e0caa1e1fb0ddeb7cafd5/mlx_lm/models/glm4_moe_lite.py" } ] }, { "id": "microsoft/Phi-3-vision-128k-instruct", "name": "Phi-3-vision-128k-instruct BF16", "short_name": "Phi 3 vision 128k", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 250174, "tags": [ "transformers", "safetensors", "phi3_v", "text-generation", "nlp", "code", "vision", "conversational", "custom_code", "multilingual", "license:mit", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "phi3_v (Phi3VForCausalLM)", "total_params_b": 4.14662144, "active_params_b": 3.722578944, "layers": 32, "attention_heads": 32, "kv_heads": 32, "head_dim": 96, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 8.29324288, "active_weight_gb": 7.445157888, "kv_alloc_gb_per_1k_tokens": 0.393216, "kv_read_gb_per_1k_tokens": 0.393216, "notes": "Manually audited on 2026-07-06 from HF API commit ed2772fabe9dc9acd0caad54b62761d92520cc44, config.json, remote modeling code, model card, safetensors index, and range-read shard headers. The authoritative bounds profile uses exact BF16 stored bytes: 8.29324288 GB resident, 7.445157888 GB swept text/logit traffic, 0.848084992 GB resident-only vision plus input-embedding tensors, and full-context BF16 MHA KV." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/microsoft/Phi-3-vision-128k-instruct" }, { "label": "Model config", "url": "https://huggingface.co/microsoft/Phi-3-vision-128k-instruct/raw/ed2772fabe9dc9acd0caad54b62761d92520cc44/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/microsoft/Phi-3-vision-128k-instruct" }, { "label": "Safetensors index", "url": "https://huggingface.co/microsoft/Phi-3-vision-128k-instruct/raw/ed2772fabe9dc9acd0caad54b62761d92520cc44/model.safetensors.index.json" }, { "label": "Remote modeling code", "url": "https://huggingface.co/microsoft/Phi-3-vision-128k-instruct/raw/ed2772fabe9dc9acd0caad54b62761d92520cc44/modeling_phi3_v.py" } ] }, { "id": "RedHatAI/diffusiongemma-26B-A4B-it-FP8-dynamic", "name": "diffusiongemma-26B-A4B-it-FP8-dynamic FP8", "short_name": "diffusiongemma 26B A4B it FP8 dynamic", "license": "unknown", "hf_pipeline_tag": "unknown", "hf_downloads": 277204, "tags": [ "safetensors", "diffusion_gemma", "fp8", "vllm", "compressed-tensors", "llm-compressor", "base_model:google/diffusiongemma-26B-A4B-it", "base_model:quantized:google/diffusiongemma-26B-A4B-it", "region:us" ], "architecture": { "type": "Block diffusion MoE transformer", "detail": "diffusion_gemma (DiffusionGemmaForBlockDiffusion)", "total_params_b": 25.840547948, "active_params_b": 4, "layers": 30, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "FP8", "weight_bytes_per_param": 1, "resident_weight_gb": 27.197665496, "active_weight_gb": 4.611052604, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manual fail-closed correction on 2026-07-06 from current HF API commit 9133fd54d9610b569b17341ca7e24efb9687ef98, pinned config, recipe, generation config, model card, safetensors index, and direct range-read header. The exact compressed-tensors resident package is 27.197665496 GB. Production bounds are disabled by the unsupported profile because DiffusionGemma uses block diffusion over a 256-token canvas, which Bounds Engine v1 does not model as ordinary autoregressive decode." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/RedHatAI/diffusiongemma-26B-A4B-it-FP8-dynamic" }, { "label": "Model config", "url": "https://huggingface.co/RedHatAI/diffusiongemma-26B-A4B-it-FP8-dynamic/raw/9133fd54d9610b569b17341ca7e24efb9687ef98/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/diffusiongemma-26B-A4B-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/RedHatAI/diffusiongemma-26B-A4B-it-FP8-dynamic" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/RedHatAI/diffusiongemma-26B-A4B-it-FP8-dynamic/raw/9133fd54d9610b569b17341ca7e24efb9687ef98/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen2.5-Coder-3B-Instruct", "name": "Qwen2.5-Coder-3B-Instruct BF16", "short_name": "Qwen2.5 Coder 3B", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 226000, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "code", "codeqwen", "chat", "qwen", "qwen-coder", "conversational", "en", "arxiv:2409.12186", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-Coder-3B", "base_model:finetune:Qwen/Qwen2.5-Coder-3B", "license:other", "text-generation-inference", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 3.085938688, "active_params_b": 3.085938688, "layers": 36, "attention_heads": 16, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 6.171877376, "active_weight_gb": 6.171877376, "kv_alloc_gb_per_1k_tokens": 0.036864, "kv_read_gb_per_1k_tokens": 0.036864, "notes": "Manually audited on 2026-07-06 from HF API commit 488639f1ff808d1d3d0ba301aef8c11461451ec5, config.json, base config comparison, model card, safetensors index, and range-read shard headers. The exact BF16 resident and swept text-decode footprint is 6.171877376 GB with full-context BF16 MHA KV." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-3B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-3B-Instruct/raw/488639f1ff808d1d3d0ba301aef8c11461451ec5/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-3B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Coder-3B-Instruct" }, { "label": "Safetensors index", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-3B-Instruct/raw/488639f1ff808d1d3d0ba301aef8c11461451ec5/model.safetensors.index.json" } ] }, { "id": "TheBloke/Mistral-7B-Instruct-v0.2-AWQ", "name": "Mistral-7B-Instruct-v0.2-AWQ 4-bit", "short_name": "Mistral 7B Instruct v0.2", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 275123, "tags": [ "transformers", "safetensors", "mistral", "text-generation", "finetuned", "conversational", "arxiv:2310.06825", "base_model:mistralai/Mistral-7B-Instruct-v0.2", "base_model:quantized:mistralai/Mistral-7B-Instruct-v0.2", "license:apache-2.0", "text-generation-inference", "4-bit", "awq", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "mistral (MistralForCausalLM)", "total_params_b": 7.241732096, "active_params_b": 7.110660096, "layers": 32, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "AWQ 4-bit + F16 side tensors", "weight_bytes_per_param": 0.573177277614662, "resident_weight_gb": 4.150796288, "active_weight_gb": 3.888652288, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, model card, pinned AWQ config, quant_config, current base config comparison, linked-object HEAD check, and direct range-read safetensors header. Exact tensor payload bytes: 4.150796288 GB resident, 3.888652288 GB swept ordinary text traffic, and 0.262144000 GB resident-only input embedding bytes. The safetensors linked object has 83944 bytes of header overhead outside tensor payload." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/TheBloke/Mistral-7B-Instruct-v0.2-AWQ" }, { "label": "Model config", "url": "https://huggingface.co/TheBloke/Mistral-7B-Instruct-v0.2-AWQ/raw/f970a2bb89d5c2f9d217dc337f39e24625d6462a/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/TheBloke/Mistral-7B-Instruct-v0.2-AWQ" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/TheBloke/Mistral-7B-Instruct-v0.2-AWQ/resolve/f970a2bb89d5c2f9d217dc337f39e24625d6462a/model.safetensors" } ] }, { "id": "Qwen/Qwen2.5-1.5B-Instruct-GGUF", "name": "Qwen2.5-1.5B-Instruct-GGUF GGUF", "short_name": "Qwen2.5 1.5B Instruct", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 234624, "tags": [ "gguf", "chat", "text-generation", "en", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-1.5B-Instruct", "base_model:quantized:Qwen/Qwen2.5-1.5B-Instruct", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 GGUF selected FP16 artifact", "total_params_b": 1.777088, "active_params_b": 1.543714304, "layers": 28, "attention_heads": 12, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 8192, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF FP16/F32 selected file", "weight_bytes_per_param": 2, "resident_weight_gb": 3.560416288, "active_weight_gb": 3.0877184, "kv_alloc_gb_per_1k_tokens": 0.028672, "kv_read_gb_per_1k_tokens": 0.028672, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, model card, audited BF16 base profile comparison, linked GGUF file sizes, and direct selected FP16 GGUF tensor-index range read. The authoritative bounds profile uses qwen2.5-1.5b-instruct-fp16.gguf because it matches API gguf.totalFileSize: 3.560416288 GB resident and 3.087718400 GB ordinary text-decode swept traffic after excluding token_embd.weight. The selected GGUF header records 8192 context, while the card and BF16 base record 32768 full context with 8192 generation." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-1.5B-Instruct-GGUF" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct-GGUF/resolve/91cad51170dc346986eccefdc2dd33a9da36ead9/qwen2.5-1.5b-instruct-fp16.gguf" } ] }, { "id": "douyamv/Gemma-4-31B-JANG_4M-CRACK-GGUF", "name": "Gemma-4-31B-JANG_4M-CRACK-GGUF GGUF", "short_name": "Gemma 4 31B JANG 4M CRACK", "license": "gemma", "hf_pipeline_tag": "text-generation", "hf_downloads": 298877, "tags": [ "gguf", "gemma4", "quantized", "31b", "text-generation", "en", "license:gemma", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text GGUF Q4_K_M selected artifact", "total_params_b": 30.697345596, "active_params_b": 30.697345596, "layers": 60, "attention_heads": 32, "kv_heads": 16, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q4_K_M selected file", "weight_bytes_per_param": 0.6087515705734181, "resident_weight_gb": 18.687057344, "active_weight_gb": 18.671229168, "kv_alloc_gb_per_1k_tokens": 0.028868608, "kv_read_gb_per_1k_tokens": 0.0466944, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, model card, audited Google Gemma 4 31B IT profile comparison, linked GGUF file sizes, and direct selected Q4_K_M GGUF tensor-index range read. The model card labels Q4_K_M as the best size/quality balance and uses it in the llama.cpp example; HF API gguf.totalFileSize points at Q3_K_M, so that mismatch is recorded in the authoritative profile. Bound weights use 18.687057344 GB resident file size and 18.671229168 GB ordinary swept tensor spans. Layered Gemma 4 KV supersedes the generated full-context estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/douyamv/Gemma-4-31B-JANG_4M-CRACK-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/google/gemma-4-31B-it/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-31B-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/douyamv/Gemma-4-31B-JANG_4M-CRACK-GGUF" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/douyamv/Gemma-4-31B-JANG_4M-CRACK-GGUF/resolve/ed3a8120d4da414573688d0130f061d3b4130372/gemma-4-31b-jang-crack-Q4_K_M.gguf" } ] }, { "id": "unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF", "name": "Qwen3-Coder-30B-A3B-Instruct-GGUF GGUF", "short_name": "Qwen3 Coder 30B A3B Instruct", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 239510, "tags": [ "transformers", "gguf", "unsloth", "qwen3", "qwen", "text-generation", "arxiv:2505.09388", "base_model:Qwen/Qwen3-Coder-30B-A3B-Instruct", "base_model:quantized:Qwen/Qwen3-Coder-30B-A3B-Instruct", "license:apache-2.0", "endpoints_compatible", "region:us", "imatrix", "conversational" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_moe (Qwen3MoeForCausalLM)", "total_params_b": 30.532122624, "active_params_b": 3.041867776, "layers": 48, "attention_heads": 32, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "GGUF IQ4_NL", "weight_bytes_per_param": 0.5669695777519487, "resident_weight_gb": 17.310784672, "active_weight_gb": 1.841543168, "expert_param_b": 0.226492416, "fixed_param_b": 1.229928448, "shared_expert_param_b": 0, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, model card, audited BF16 base profile/config, selected linked-object metadata, sibling HEAD checks, and a direct GGUF header range read. The API-selected artifact is Qwen3-Coder-30B-A3B-Instruct-IQ4_NL.gguf at 17.310784672 GB resident. Ordinary single-session MoE text-decode traffic is 1.841543168 GB: 0.822327296 GB fixed traffic plus 8 routed expert groups at 0.127401984 GB each. The audited profile supersedes the generated BF16-derived estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/resolve/b17cb02dd882d5b6ab62fc777ad2995f19668350/Qwen3-Coder-30B-A3B-Instruct-IQ4_NL.gguf" } ] }, { "id": "RedHatAI/Meta-Llama-3.1-8B-FP8", "name": "Meta-Llama-3.1-8B-FP8 FP8", "short_name": "Meta Llama 3.1 8B FP8", "license": "llama3.1", "hf_pipeline_tag": "text-generation", "hf_downloads": 238957, "tags": [ "transformers", "safetensors", "llama", "text-generation", "fp8", "vllm", "en", "de", "fr", "it", "pt", "hi", "es", "th", "base_model:meta-llama/Llama-3.1-8B", "base_model:quantized:meta-llama/Llama-3.1-8B", "license:llama3.1", "text-generation-inference", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "llama (LlamaForCausalLM)", "total_params_b": 8.030261696, "active_params_b": 7.50492512, "layers": 32, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "FP8", "weight_bytes_per_param": 1, "resident_weight_gb": 9.081201536, "active_weight_gb": 8.030528384, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "notes": "Manually audited on 2026-07-06 from current Hugging Face 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 audited RedHatAI Instruct FP8 sibling. Bound weights separate the 9.081201536 GB resident compressed-tensors package from the 8.030528384 GB ordinary text-decode swept subset; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/RedHatAI/Meta-Llama-3.1-8B-FP8" }, { "label": "Model config", "url": "https://huggingface.co/RedHatAI/Meta-Llama-3.1-8B-FP8/raw/881d74b74ad8ebe8775fb178b1fd0c935477f1bc/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/meta-llama/Llama-3.1-8B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/RedHatAI/Meta-Llama-3.1-8B-FP8" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/RedHatAI/Meta-Llama-3.1-8B-FP8/raw/881d74b74ad8ebe8775fb178b1fd0c935477f1bc/model.safetensors.index.json" } ] }, { "id": "meta-llama/Llama-Guard-3-8B", "name": "Llama-Guard-3-8B BF16", "short_name": "Llama Guard 3 8B", "license": "llama3.1", "hf_pipeline_tag": "text-generation", "hf_downloads": 244402, "tags": [ "transformers", "safetensors", "llama", "text-generation", "facebook", "meta", "pytorch", "llama-3", "conversational", "en", "arxiv:2407.21783", "arxiv:2312.06674", "arxiv:2204.05862", "arxiv:2308.01263", "base_model:meta-llama/Llama-3.1-8B", "base_model:finetune:meta-llama/Llama-3.1-8B", "license:llama3.1", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gated config unavailable", "total_params_b": 8.030261248, "active_params_b": 8.030261248, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 16.060522496, "active_weight_gb": 16.060522496, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata and gated access checks. The API reports BF16 safetensors parameters, but raw config, README, original/params.json, safetensors index, and hf download with the configured osolmaz CLI token were denied because the repository requires approval. Production bounds use the fail-closed unsupported profile until config and tensor evidence are directly accessible." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/meta-llama/Llama-Guard-3-8B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-Guard-3-8B" }, { "label": "Gated config and tensor access check", "url": "https://huggingface.co/meta-llama/Llama-Guard-3-8B/raw/7327bd9f6efbbe6101dc6cc4736302b3cbb6e425/config.json" } ] }, { "id": "microsoft/Phi-3-mini-128k-instruct", "name": "Phi-3-mini-128k-instruct BF16", "short_name": "Phi 3 mini 128k", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 243935, "tags": [ "transformers", "safetensors", "phi3", "text-generation", "nlp", "code", "conversational", "custom_code", "en", "license:mit", "eval-results", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "phi3 (Phi3ForCausalLM)", "total_params_b": 3.821079552, "active_params_b": 3.722578944, "layers": 32, "attention_heads": 32, "kv_heads": 32, "head_dim": 96, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 7.642159104, "active_weight_gb": 7.445157888, "kv_alloc_gb_per_1k_tokens": 0.393216, "kv_read_gb_per_1k_tokens": 0.393216, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, model card, pinned config, pinned remote modeling code, safetensors index, direct shard header range reads, and comparison against the audited Phi-3.5 Mini Instruct profile. Bound weights separate the 7.642159104 GB resident BF16 checkpoint from the 7.445157888 GB ordinary text-decode swept subset; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/microsoft/Phi-3-mini-128k-instruct" }, { "label": "Model config", "url": "https://huggingface.co/microsoft/Phi-3-mini-128k-instruct/raw/f3c06aed622e14ca0abf5115094e4fc9a9948f36/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/microsoft/Phi-3-mini-128k-instruct" }, { "label": "Pinned remote modeling code", "url": "https://huggingface.co/microsoft/Phi-3-mini-128k-instruct/raw/f3c06aed622e14ca0abf5115094e4fc9a9948f36/modeling_phi3.py" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/microsoft/Phi-3-mini-128k-instruct/resolve/f3c06aed622e14ca0abf5115094e4fc9a9948f36/model.safetensors.index.json" } ] }, { "id": "lmstudio-community/gemma-4-E2B-it-MLX-4bit", "name": "gemma-4-E2B-it-MLX-4bit 4-bit", "short_name": "gemma 4 E2B it MLX 4bit", "license": "apache-2.0", "hf_pipeline_tag": "any-to-any", "hf_downloads": 242033, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "mlx", "any-to-any", "base_model:google/gemma-4-E2B-it", "base_model:quantized:google/gemma-4-E2B-it", "license:apache-2.0", "endpoints_compatible", "4-bit", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 5.123178979, "active_params_b": 5.123178979, "layers": 35, "attention_heads": 8, "kv_heads": 1, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "MLX 4-bit affine", "weight_bytes_per_param": 0.8470422649272819, "resident_weight_gb": 4.339549126, "active_weight_gb": 2.071972422, "kv_alloc_gb_per_1k_tokens": 0.03584, "kv_read_gb_per_1k_tokens": 0.03584, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, served MLX config, audited Google config comparison, model card, safetensors index, and direct header range read. Bound weights separate the exact 4.339549126 GB resident MLX package from the 2.071972422 GB ordinary text-decode swept traffic; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/gemma-4-E2B-it-MLX-4bit" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-E2B-it-MLX-4bit" }, { "label": "Pinned MLX model config", "url": "https://huggingface.co/lmstudio-community/gemma-4-E2B-it-MLX-4bit/raw/34f17a2e0a341dcbacd9b90ee604521ac88defc7/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-E2B-it" }, { "label": "Pinned Google base config", "url": "https://huggingface.co/google/gemma-4-E2B-it/raw/70af34e20bd4b7a91f0de6b22675850c43922a03/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/lmstudio-community/gemma-4-E2B-it-MLX-4bit/raw/34f17a2e0a341dcbacd9b90ee604521ac88defc7/model.safetensors.index.json" } ] }, { "id": "unsloth/gpt-oss-120b-GGUF", "name": "gpt-oss-120b-GGUF GGUF", "short_name": "gpt oss 120b", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 184379, "tags": [ "transformers", "gguf", "openai", "unsloth", "text-generation", "base_model:openai/gpt-oss-120b", "base_model:quantized:openai/gpt-oss-120b", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "MoE transformer", "detail": "gpt_oss (GptOssForCausalLM)", "total_params_b": 116.829156672, "active_params_b": 5.1, "layers": 36, "attention_heads": 64, "kv_heads": 8, "head_dim": 64, "max_context_tokens": 131072, "routed_experts": 128, "routed_experts_per_token": 4, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "GGUF F16/MXFP4", "weight_bytes_per_param": 0.5595265735892001, "resident_weight_gb": 65.369017728, "active_weight_gb": 5.032942848, "kv_alloc_gb_per_1k_tokens": 0.073728, "kv_read_gb_per_1k_tokens": 0.073728, "notes": "Manually corrected on 2026-07-06 from current Hugging Face API metadata, model card, served config, OpenAI base config comparison, selected linked-object HEAD checks, and a direct GGUF header/tensor-index range read. The authoritative profile targets the API-selected gpt-oss-120b-F16.gguf artifact: 65.369017728 GB resident file size, 64.197727488 GB main resident tensor traffic, 1.171290240 GB resident-only input embedding plus GGUF overhead, and 5.032942848 GB single-session ordinary text-decode swept traffic." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/gpt-oss-120b-GGUF" }, { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/gpt-oss-120b-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/unsloth/gpt-oss-120b-GGUF/raw/ff1a82da6ad466e32284fa3d2b86694db3204789/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/openai/gpt-oss-120b" }, { "label": "Pinned OpenAI base config", "url": "https://huggingface.co/openai/gpt-oss-120b/raw/b5c939de8f754692c1647ca79fbf85e8c1e70f8a/config.json" }, { "label": "Selected F16 GGUF artifact", "url": "https://huggingface.co/unsloth/gpt-oss-120b-GGUF/resolve/ff1a82da6ad466e32284fa3d2b86694db3204789/gpt-oss-120b-F16.gguf" } ] }, { "id": "google/paligemma-3b-ft-cococap-448", "name": "paligemma-3b-ft-cococap-448 F32", "short_name": "paligemma 3b ft cococap 448", "license": "gemma", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 231962, "tags": [ "transformers", "safetensors", "paligemma", "image-text-to-text", "arxiv:2310.09199", "arxiv:2303.15343", "arxiv:2403.08295", "arxiv:1706.03762", "arxiv:2010.11929", "arxiv:2209.06794", "arxiv:2209.04372", "arxiv:2103.01913", "arxiv:2401.06209", "arxiv:2305.10355", "arxiv:2205.12522", "arxiv:2110.11624", "arxiv:2108.03353", "arxiv:2010.04295", "arxiv:2203.10244", "arxiv:1810.12440", "arxiv:1905.13648", "arxiv:1608.00272", "arxiv:1908.04913", "arxiv:2407.07726", "license:gemma", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "PaliGemmaForConditionalGeneration (gated config unavailable)", "total_params_b": 2.924351216, "active_params_b": 2.924351216, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "F32", "weight_bytes_per_param": 4, "resident_weight_gb": 11.697404864, "active_weight_gb": 11.697404864, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manual fail-closed correction on 2026-07-06 from current Hugging Face API metadata and gated access checks. The API exposes F32 parameter metadata but raw README, config, safetensors index, and shard bytes require manual approval, so production bounds are disabled by the unsupported profile rather than inferring PaliGemma KV geometry." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/paligemma-3b-ft-cococap-448" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/paligemma-3b-ft-cococap-448" }, { "label": "Gated config access check", "url": "https://huggingface.co/google/paligemma-3b-ft-cococap-448/raw/cb5ff322983f2ebcd916c3394ffcef78ed4e9a9b/config.json" } ] }, { "id": "RedHatAI/DeepSeek-Coder-V2-Lite-Instruct-FP8", "name": "DeepSeek-Coder-V2-Lite-Instruct-FP8 FP8", "short_name": "DeepSeek Coder V2 Lite Instruct FP8", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 245921, "tags": [ "transformers", "safetensors", "deepseek_v2", "text-generation", "fp8", "vllm", "conversational", "custom_code", "license:other", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "deepseek_v2 (DeepseekV2ForCausalLM)", "total_params_b": 15.706484224, "active_params_b": 2.451435008, "layers": 27, "attention_heads": 16, "kv_heads": 16, "head_dim": 128, "max_context_tokens": 163840, "routed_experts": 64, "routed_experts_per_token": 6, "shared_experts_per_token": 2 }, "adapter": { "kind": "moe", "weight_precision": "FP8", "weight_bytes_per_param": 1, "resident_weight_gb": 16.129490408, "active_weight_gb": 2.6646898, "kv_alloc_gb_per_1k_tokens": 0.27648, "kv_read_gb_per_1k_tokens": 0.27648, "notes": "Manually corrected on 2026-07-06 from current Hugging Face API metadata, model card, pinned static FP8 config, audited BF16 base config comparison, safetensors index, and direct range-read shard headers. Bound weights separate the 16.129490408 GB resident FP8/BF16/F32 package from 2.664689800 GB single-token active ordinary decode traffic; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/RedHatAI/DeepSeek-Coder-V2-Lite-Instruct-FP8" }, { "label": "Model config", "url": "https://huggingface.co/RedHatAI/DeepSeek-Coder-V2-Lite-Instruct-FP8/raw/efe1ced428db63e7ccbcf367334596f77e9af140/config.json" }, { "label": "Base model config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct/raw/e434a23f91ba5b4923cf6c9d9a238eb4a08e3a11/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/RedHatAI/DeepSeek-Coder-V2-Lite-Instruct-FP8" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/RedHatAI/DeepSeek-Coder-V2-Lite-Instruct-FP8/resolve/efe1ced428db63e7ccbcf367334596f77e9af140/model.safetensors.index.json" } ] }, { "id": "prithivMLmods/gemma-4-E4B-it-NVFP4", "name": "gemma-4-E4B-it-NVFP4 NVFP4", "short_name": "gemma 4 E4B it NVFP4", "license": "apache-2.0", "hf_pipeline_tag": "any-to-any", "hf_downloads": 243514, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "text-generation-inference", "vllm", "any-to-any", "en", "base_model:google/gemma-4-E4B-it", "base_model:quantized:google/gemma-4-E4B-it", "license:apache-2.0", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 6.905065317, "active_params_b": 2.937316005, "layers": 42, "attention_heads": 8, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "NVFP4", "weight_bytes_per_param": 1.6718842504759466, "resident_weight_gb": 11.544469952, "active_weight_gb": 3.608971328, "kv_alloc_gb_per_1k_tokens": 0.028672, "kv_read_gb_per_1k_tokens": 0.028672, "kv_alloc_fixed_gb": 0.03670016, "kv_read_fixed_gb": 0.03670016, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, model card, served compressed-tensors config, recipe.yaml, base Google config comparison, linked-object tree size, and direct safetensors header range read. Bound weights separate the exact 11.544469952 GB resident NVFP4/U8/F8/BF16/F32 package from the 3.608971328 GB ordinary text-decode swept subset; hybrid Gemma 4 E4B KV uses full-context plus fixed sliding-window cache terms. The audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/prithivMLmods/gemma-4-E4B-it-NVFP4" }, { "label": "Model config", "url": "https://huggingface.co/prithivMLmods/gemma-4-E4B-it-NVFP4/raw/bd4ca527287ab7de27f32ec80fa6235d7c4b08e0/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-E4B-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/prithivMLmods/gemma-4-E4B-it-NVFP4" }, { "label": "Pinned Google Gemma 4 E4B IT base config", "url": "https://huggingface.co/google/gemma-4-E4B-it/raw/fee6332c1abaafb77f6f9624236c63aa2f1d0187/config.json" }, { "label": "Quantization recipe", "url": "https://huggingface.co/prithivMLmods/gemma-4-E4B-it-NVFP4/raw/bd4ca527287ab7de27f32ec80fa6235d7c4b08e0/recipe.yaml" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/prithivMLmods/gemma-4-E4B-it-NVFP4/resolve/bd4ca527287ab7de27f32ec80fa6235d7c4b08e0/model.safetensors" } ] }, { "id": "Qwen/Qwen2.5-Coder-7B-Instruct-AWQ", "name": "Qwen2.5-Coder-7B-Instruct-AWQ 4-bit", "short_name": "Qwen2.5 Coder 7B Instruct", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 258690, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "code", "codeqwen", "chat", "qwen", "qwen-coder", "conversational", "en", "arxiv:2409.12186", "arxiv:2309.00071", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-Coder-7B-Instruct", "base_model:quantized:Qwen/Qwen2.5-Coder-7B-Instruct", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "4-bit", "awq", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 7.615616512, "active_params_b": 7.070619136, "layers": 28, "attention_heads": 28, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "4-bit", "weight_bytes_per_param": 0.5, "resident_weight_gb": 5.570747392, "active_weight_gb": 4.48075264, "kv_alloc_gb_per_1k_tokens": 0.057344, "kv_read_gb_per_1k_tokens": 0.057344, "notes": "Manually audited on 2026-07-06 from HF API commit 8e8ed243bbe6f9a5aff549a0924562fc719b2b8a, served AWQ config, model card, audited BF16 base profile comparison, safetensors index, linked-object size checks, and direct safetensors header range reads. The authoritative bounds profile uses exact tensor payload bytes: 5.570747392 GB resident, 4.480752640 GB swept ordinary text traffic, and 1.089994752 GB resident-only input embedding." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct-AWQ" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct-AWQ/raw/8e8ed243bbe6f9a5aff549a0924562fc719b2b8a/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Coder-7B-Instruct-AWQ" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct-AWQ/raw/8e8ed243bbe6f9a5aff549a0924562fc719b2b8a/model.safetensors.index.json" } ] }, { "id": "MaziyarPanahi/Phi-3.5-mini-instruct-GGUF", "name": "Phi-3.5-mini-instruct-GGUF GGUF Q4_K_M", "short_name": "Phi 3.5 mini instruct", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 210087, "tags": [ "gguf", "mistral", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:microsoft/Phi-3.5-mini-instruct", "base_model:quantized:microsoft/Phi-3.5-mini-instruct", "license:mit", "region:us", "imatrix", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "phi3 (Phi3ForCausalLM, Q4_K_M GGUF)", "total_params_b": 3.821079648, "active_params_b": 3.72257904, "layers": 32, "attention_heads": 32, "kv_heads": 32, "head_dim": 96, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q4_K_M", "weight_bytes_per_param": 0.6263236646356344, "resident_weight_gb": 2.393232608, "active_weight_gb": 2.33708736, "kv_alloc_gb_per_1k_tokens": 0.393216, "kv_read_gb_per_1k_tokens": 0.393216, "notes": "Manually audited on 2026-07-06 from HF API commit 53fbed70b1fbf95c833c095523825cce68ebc063, model card, audited Microsoft BF16 base profile, linked GGUF file size checks, stale repo-config check, and a direct selected Q4_K_M GGUF header range read. The authoritative bounds profile uses the selected Q4_K_M artifact: 2.393232608 GB resident, 2.337087360 GB swept text traffic, 0.056145248 GB resident-only token embedding plus GGUF metadata/header/file overhead, 131072-token context, and full-context FP16 K/V cache. API gguf.totalFileSize selects the smaller IQ1_M file, but rendered runtime examples select Q4_K_M." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/Phi-3.5-mini-instruct-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Phi-3.5-mini-instruct-GGUF" }, { "label": "Base model card", "url": "https://huggingface.co/microsoft/Phi-3.5-mini-instruct" }, { "label": "Stale repo config", "url": "https://huggingface.co/MaziyarPanahi/Phi-3.5-mini-instruct-GGUF/raw/53fbed70b1fbf95c833c095523825cce68ebc063/config.json" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/MaziyarPanahi/Phi-3.5-mini-instruct-GGUF/resolve/53fbed70b1fbf95c833c095523825cce68ebc063/Phi-3.5-mini-instruct.Q4_K_M.gguf" } ] }, { "id": "VlSav/Vikhr-Nemo-12B-Instruct-R-21-09-24-Q8_0-GGUF", "name": "Vikhr-Nemo-12B-Instruct-R-21-09-24-Q8_0-GGUF GGUF 8-bit", "short_name": "Vikhr Nemo 12B Instruct R 21 09 24 Q8 0", "license": "apache-2.0", "hf_pipeline_tag": "unknown", "hf_downloads": 199862, "tags": [ "gguf", "llama-cpp", "gguf-my-repo", "en", "ru", "dataset:Vikhrmodels/GrandMaster-PRO-MAX", "dataset:Vikhrmodels/Grounded-RAG-RU-v2", "base_model:Vikhrmodels/Vikhr-Nemo-12B-Instruct-R-21-09-24", "base_model:quantized:Vikhrmodels/Vikhr-Nemo-12B-Instruct-R-21-09-24", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "mistral/llama GGUF (MistralForCausalLM, Q8_0 GGUF)", "total_params_b": 12.24780288, "active_params_b": 11.576704, "layers": 40, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 1024000, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q8_0", "weight_bytes_per_param": 1.0632430295938924, "resident_weight_gb": 13.02239104, "active_weight_gb": 12.30146624, "kv_alloc_gb_per_1k_tokens": 0.16384, "kv_read_gb_per_1k_tokens": 0.16384, "notes": "Manually audited on 2026-07-06 from HF API commit 6a36ecf1d5586f8ce0d4a29d4c95ef2f3ad7f0e1, model card, original Vikhr API/config metadata, Mistral Nemo base config comparison, linked GGUF file size check, and a direct selected Q8_0 GGUF header range read. The authoritative bounds profile uses the single selected Q8_0 artifact: 13.022391040 GB resident, 12.301466240 GB swept text traffic, 0.720924800 GB resident-only token embedding plus GGUF metadata/header/file overhead, 1024000-token context, and full-context FP16 K/V cache." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/VlSav/Vikhr-Nemo-12B-Instruct-R-21-09-24-Q8_0-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/VlSav/Vikhr-Nemo-12B-Instruct-R-21-09-24-Q8_0-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/Vikhrmodels/Vikhr-Nemo-12B-Instruct-R-21-09-24/raw/7499757d9f41a2965b0e9db94c976e0982e292b8/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Vikhrmodels/Vikhr-Nemo-12B-Instruct-R-21-09-24" }, { "label": "Selected GGUF artifact", "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" } ] }, { "id": "nvidia/Llama-3_3-Nemotron-Super-49B-v1_5-FP8", "name": "Llama-3_3-Nemotron-Super-49B-v1_5-FP8 FP8", "short_name": "Llama 3 3 Nemotron Super 49B v1 5 FP8", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 233019, "tags": [ "transformers", "safetensors", "nemotron-nas", "text-generation", "nvidia", "llama-3", "pytorch", "conversational", "custom_code", "en", "arxiv:2411.19146", "arxiv:2505.00949", "arxiv:2502.00203", "license:other", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "nemotron-nas (DeciLMForCausalLM)", "total_params_b": 49.867146186, "active_params_b": 48.816473034, "layers": 80, "attention_heads": 64, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "FP8 F8_E4M3 + BF16/F32 side tensors", "weight_bytes_per_param": 1, "resident_weight_gb": 51.96956036, "active_weight_gb": 49.868214056, "kv_alloc_gb_per_1k_tokens": 0.100352, "kv_read_gb_per_1k_tokens": 0.100352, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, pinned served DeciLM config, audited BF16 DeciLM runtime comparison, ModelOpt quantization config, block config dataclasses, safetensors index, and direct safetensors shard header range reads. downloads=233019. Exact mixed ModelOpt tensor payload bytes: 51.969560360 GB resident, 49.868214056 GB swept ordinary text traffic, and 2.101346304 GB resident-only input embedding. hf_quant_config.json records FP8 KV cache, so KV traffic is charged at one byte per scalar. The audited DeciLM profile supersedes the generated estimate, which incorrectly charged 64 KV heads across all 80 layers and missed BF16/F32 side tensors." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1_5-FP8" }, { "label": "Model config", "url": "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1_5-FP8/raw/04822723e77e036ddf2d24e83c6d469d3b009252/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/Llama-3_3-Nemotron-Super-49B-v1_5-FP8" }, { "label": "ModelOpt quantization config", "url": "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1_5-FP8/raw/04822723e77e036ddf2d24e83c6d469d3b009252/hf_quant_config.json" }, { "label": "Audited BF16 DeciLM runtime", "url": "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1_5/raw/420ba7d28211abf116b8b103ab700d92619daf98/modeling_decilm.py" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1_5-FP8/raw/04822723e77e036ddf2d24e83c6d469d3b009252/model.safetensors.index.json" } ] }, { "id": "mlx-community/Kimi-K2.5", "name": "Kimi-K2.5 4-bit", "short_name": "Kimi K2.5", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 225434, "tags": [ "mlx", "safetensors", "kimi_k25", "text-generation", "conversational", "custom_code", "base_model:moonshotai/Kimi-K2.5", "base_model:quantized:moonshotai/Kimi-K2.5", "license:other", "4-bit", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "kimi_k25 (KimiK25ForConditionalGeneration / DeepseekV3ForCausalLM)", "total_params_b": 1058.589420528, "active_params_b": 32, "layers": 61, "attention_heads": 64, "kv_heads": 64, "head_dim": 192, "max_context_tokens": 262144, "routed_experts": 384, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "MLX 4-bit affine + BF16/F32 side tensors", "weight_bytes_per_param": 0.6407033845077003, "resident_weight_gb": 657.623228416, "active_weight_gb": 34.307710976, "kv_alloc_gb_per_1k_tokens": 2.49856, "kv_read_gb_per_1k_tokens": 2.49856, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, MLX model card, pinned served config, audited Moonshot Kimi K2.5 base config comparison, custom runtime review, safetensors index, and direct safetensors shard header range reads. downloads=225434. Bound weights separate the exact 657.623228416 GB resident MLX package from 34.307710976 GB single-session ordinary text-decode MoE traffic; the package contains no vision_tower or mm_projector tensors, and the audited profile supersedes the generated ideal 4-bit estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/mlx-community/Kimi-K2.5" }, { "label": "Model config", "url": "https://huggingface.co/mlx-community/Kimi-K2.5/raw/351021afd838c866ce1a7374fce51d615773d2a8/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/moonshotai/Kimi-K2.5" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/mlx-community/Kimi-K2.5" }, { "label": "Pinned base config", "url": "https://huggingface.co/moonshotai/Kimi-K2.5/raw/4d01dfe0332d63057c186e0b262165819efb6611/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/mlx-community/Kimi-K2.5/raw/351021afd838c866ce1a7374fce51d615773d2a8/model.safetensors.index.json" } ] }, { "id": "lmstudio-community/gemma-4-E2B-it-MLX-8bit", "name": "gemma-4-E2B-it-MLX-8bit 8-bit", "short_name": "gemma 4 E2B it MLX 8bit", "license": "apache-2.0", "hf_pipeline_tag": "any-to-any", "hf_downloads": 231439, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "mlx", "any-to-any", "base_model:google/gemma-4-E2B-it", "base_model:quantized:google/gemma-4-E2B-it", "license:apache-2.0", "endpoints_compatible", "8-bit", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 5.123178979, "active_params_b": 5.123178979, "layers": 35, "attention_heads": 8, "kv_heads": 1, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "MLX 8-bit affine + BF16 side tensors", "weight_bytes_per_param": 1.1489604056637819, "resident_weight_gb": 5.886329798, "active_weight_gb": 2.589379142, "kv_alloc_gb_per_1k_tokens": 0.006144, "kv_read_gb_per_1k_tokens": 0.014336, "kv_alloc_fixed_gb": 0.006291456, "kv_read_fixed_gb": 0.014680064, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, served MLX config, audited Google config comparison, model card, safetensors index, and direct two-shard header range reads. downloads=231439. Bound weights separate the exact 5.886329798 GB resident MLX package from the 2.589379142 GB ordinary text-decode swept traffic; hybrid Gemma 4 E2B KV uses full-context plus fixed sliding-window terms. The audited profile supersedes the generated ideal 8-bit estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/gemma-4-E2B-it-MLX-8bit" }, { "label": "Model config", "url": "https://huggingface.co/lmstudio-community/gemma-4-E2B-it-MLX-8bit/raw/df2b6aabee32f10605d5333775c18ff10b9d60e5/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-E2B-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-E2B-it-MLX-8bit" }, { "label": "Pinned Google base config", "url": "https://huggingface.co/google/gemma-4-E2B-it/raw/70af34e20bd4b7a91f0de6b22675850c43922a03/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/lmstudio-community/gemma-4-E2B-it-MLX-8bit/raw/df2b6aabee32f10605d5333775c18ff10b9d60e5/model.safetensors.index.json" } ] }, { "id": "lmstudio-community/gemma-4-31B-it-QAT-GGUF", "name": "gemma-4-31B-it-QAT-GGUF GGUF 4-bit", "short_name": "gemma 4 31B it QAT", "license": "apache-2.0", "hf_pipeline_tag": "unknown", "hf_downloads": 273894, "tags": [ "gguf", "base_model:google/gemma-4-31B-it-qat-q4_0-unquantized", "base_model:quantized:google/gemma-4-31B-it-qat-q4_0-unquantized", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 30.697345596, "active_params_b": 30.697345596, "layers": 60, "attention_heads": 32, "kv_heads": 16, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q4_0 QAT", "weight_bytes_per_param": 0.5750008811934542, "resident_weight_gb": 17.651000768, "active_weight_gb": 17.635168128, "kv_alloc_gb_per_1k_tokens": 0.04096, "kv_read_gb_per_1k_tokens": 0.04096, "kv_alloc_fixed_gb": 0.8388608, "kv_read_fixed_gb": 0.8388608, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, model card, Google QAT unquantized base config, linked-object HEAD checks, and direct GGUF header/tensor-index range read. downloads=273894. The authoritative profile uses the selected gemma-4-31B-it-QAT-Q4_0.gguf artifact: 17.651000768 GB resident file size, 17.635168128 GB swept text tensor spans, 0.015832640 GB resident metadata/header/file overhead, and hybrid full-context plus fixed sliding-window FP16 KV. The separate mmproj sidecar is not included unless explicitly loaded." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/gemma-4-31B-it-QAT-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/google/gemma-4-31B-it-qat-q4_0-unquantized/raw/4f926903562062220b3e54c1385c5ef2cd40bfd1/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-31B-it-qat-q4_0-unquantized" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-31B-it-QAT-GGUF" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/lmstudio-community/gemma-4-31B-it-QAT-GGUF/resolve/5f1655eb80159b7db3f6feb9ce1a9440ab076261/gemma-4-31B-it-QAT-Q4_0.gguf" } ] }, { "id": "lmstudio-community/gemma-4-E2B-it-MLX-6bit", "name": "gemma-4-E2B-it-MLX-6bit 6-bit", "short_name": "gemma 4 E2B it MLX 6bit", "license": "apache-2.0", "hf_pipeline_tag": "any-to-any", "hf_downloads": 230149, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "mlx", "any-to-any", "base_model:google/gemma-4-E2B-it", "base_model:quantized:google/gemma-4-E2B-it", "license:apache-2.0", "endpoints_compatible", "6-bit", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 5.123178979, "active_params_b": 5.123178979, "layers": 35, "attention_heads": 8, "kv_heads": 1, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "MLX 6-bit affine + BF16 side tensors", "weight_bytes_per_param": 0.9980013352955318, "resident_weight_gb": 5.112939462, "active_weight_gb": 2.404076102, "kv_alloc_gb_per_1k_tokens": 0.006144, "kv_read_gb_per_1k_tokens": 0.014336, "kv_alloc_fixed_gb": 0.006291456, "kv_read_fixed_gb": 0.014680064, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, served MLX config, audited Google config comparison, model card, safetensors index, and direct header range read. downloads=230149. Bound weights separate the exact 5.112939462 GB resident MLX package from the 2.404076102 GB ordinary text-decode swept traffic; hybrid Gemma 4 E2B KV uses full-context plus fixed sliding-window terms. The audited profile supersedes the generated BF16/U32 estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/gemma-4-E2B-it-MLX-6bit" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-E2B-it-MLX-6bit" }, { "label": "Pinned MLX model config", "url": "https://huggingface.co/lmstudio-community/gemma-4-E2B-it-MLX-6bit/raw/182790d74c2967b8bba0a458d507cfcb0482e8a6/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-E2B-it" }, { "label": "Pinned Google base config", "url": "https://huggingface.co/google/gemma-4-E2B-it/raw/70af34e20bd4b7a91f0de6b22675850c43922a03/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/lmstudio-community/gemma-4-E2B-it-MLX-6bit/raw/182790d74c2967b8bba0a458d507cfcb0482e8a6/model.safetensors.index.json" } ] }, { "id": "lmstudio-community/gemma-4-E2B-it-MLX-5bit", "name": "gemma-4-E2B-it-MLX-5bit 5-bit", "short_name": "gemma 4 E2B it MLX 5bit", "license": "apache-2.0", "hf_pipeline_tag": "any-to-any", "hf_downloads": 230120, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "mlx", "any-to-any", "base_model:google/gemma-4-E2B-it", "base_model:quantized:google/gemma-4-E2B-it", "license:apache-2.0", "endpoints_compatible", "5-bit", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 5.123178979, "active_params_b": 5.123178979, "layers": 35, "attention_heads": 8, "kv_heads": 1, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "MLX 5-bit affine + BF16 side tensors", "weight_bytes_per_param": 0.9225218001114068, "resident_weight_gb": 4.726244294, "active_weight_gb": 2.311424582, "kv_alloc_gb_per_1k_tokens": 0.006144, "kv_read_gb_per_1k_tokens": 0.014336, "kv_alloc_fixed_gb": 0.006291456, "kv_read_fixed_gb": 0.014680064, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, served MLX config, audited Google config comparison, model card, safetensors index, and direct header range read. downloads=230120. Bound weights separate the exact 4.726244294 GB resident MLX package from the 2.311424582 GB ordinary text-decode swept traffic; hybrid Gemma 4 E2B KV uses full-context plus fixed sliding-window terms. The audited profile supersedes the generated BF16/U32 estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/gemma-4-E2B-it-MLX-5bit" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-E2B-it-MLX-5bit" }, { "label": "Pinned MLX model config", "url": "https://huggingface.co/lmstudio-community/gemma-4-E2B-it-MLX-5bit/raw/2f80e230cb30cec567f11f93835102f999423edc/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-E2B-it" }, { "label": "Pinned Google base config", "url": "https://huggingface.co/google/gemma-4-E2B-it/raw/70af34e20bd4b7a91f0de6b22675850c43922a03/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/lmstudio-community/gemma-4-E2B-it-MLX-5bit/raw/2f80e230cb30cec567f11f93835102f999423edc/model.safetensors.index.json" } ] }, { "id": "nvidia/Nemotron-Labs-Diffusion-3B-Base", "name": "Nemotron-Labs-Diffusion-3B-Base BF16", "short_name": "Nemotron Labs Diffusion 3B", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 270224, "tags": [ "transformers", "safetensors", "nemotron_labs_diffusion", "feature-extraction", "nvidia", "pytorch", "text-generation", "conversational", "custom_code", "license:other", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "nemotron_labs_diffusion (NemotronLabsDiffusionModel)", "total_params_b": 3.83165952, "active_params_b": 3.429006336, "layers": 26, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 4096, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 7.66331904, "active_weight_gb": 6.858012672, "kv_alloc_gb_per_1k_tokens": 0.106496, "kv_read_gb_per_1k_tokens": 0.106496, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, model card, pinned custom config, custom AR runtime review, and direct safetensors header range read. downloads=270224. Bound weights model only the ordinary ar_generate path: 7.663319040 GB resident BF16 checkpoint, 6.858012672 GB swept AR text-decode traffic, and full-context BF16 KV. Diffusion parallel decoding and self-speculation are outside Bounds Engine v1." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-3B-Base" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/Nemotron-Labs-Diffusion-3B-Base" }, { "label": "Pinned model config", "url": "https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-3B-Base/raw/4bd78816ba988ca8b71b8f432b00b682572b236c/config.json" }, { "label": "Pinned custom model code", "url": "https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-3B-Base/raw/4bd78816ba988ca8b71b8f432b00b682572b236c/modeling_nemotron_labs_diffusion.py" }, { "label": "Pinned custom attention code", "url": "https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-3B-Base/raw/4bd78816ba988ca8b71b8f432b00b682572b236c/modeling_ministral.py" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-3B-Base/raw/4bd78816ba988ca8b71b8f432b00b682572b236c/model.safetensors" } ] }, { "id": "VLTX/VertaLily-1.2-1B-GGUF", "name": "VertaLily-1.2-1B-GGUF GGUF", "short_name": "VertaLily 1.2 1B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 152564, "tags": [ "transformers", "gguf", "vltx", "sovereign", "techina-x", "void", "verta-lily", "text-generation", "en", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Hybrid short-conv transformer", "detail": "lfm2/vltx GGUF hybrid short-conv + attention", "total_params_b": 1.170340608, "active_params_b": 1.170340608, "layers": 16, "attention_heads": 32, "kv_heads": 8, "head_dim": 64, "max_context_tokens": 128000, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "hybrid", "weight_precision": "GGUF Q4_K_M", "weight_bytes_per_param": 0.6245142337229743, "resident_weight_gb": 0.730894368, "active_weight_gb": 0.72850944, "kv_alloc_gb_per_1k_tokens": 0.012288, "kv_read_gb_per_1k_tokens": 0.012288, "kv_alloc_fixed_gb": 0.00012288, "kv_read_fixed_gb": 0.00012288, "notes": "Manually audited on 2026-07-06 from HF API commit be658e2275a7f717e4648b9513ff5fe49bf68ff1, model card, selected linked-object HEAD checks, direct Q4_K_M GGUF header range read, and Transformers LFM2 implementation review. The authoritative profile uses the card-recommended VertaLily-1.2-1B-Q4_K_M-stable.gguf artifact: 0.730894368 GB resident, 0.728509440 GB swept tensor traffic, 0.002384928 GB GGUF overhead, six full-attention layers, and ten fixed short-conv recurrent-state layers. The HF API gguf.totalFileSize points to prop.PearlLily-unst.gguf, so the selected artifact is intentionally pinned by card recommendation." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/VLTX/VertaLily-1.2-1B-GGUF/blob/be658e2275a7f717e4648b9513ff5fe49bf68ff1/README.md" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/VLTX/VertaLily-1.2-1B-GGUF" }, { "label": "Selected Q4_K_M GGUF header audit", "url": "https://huggingface.co/VLTX/VertaLily-1.2-1B-GGUF/resolve/be658e2275a7f717e4648b9513ff5fe49bf68ff1/VertaLily-1.2-1B-Q4_K_M-stable.gguf" }, { "label": "Transformers LFM2 implementation", "url": "https://github.com/huggingface/transformers/blob/5204b4fe36956e9214b9279f1e1be2fd5dd1d9f3/src/transformers/models/lfm2/modeling_lfm2.py" } ] }, { "id": "Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-GGUF", "name": "Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-GGUF GGUF", "short_name": "Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 212029, "tags": [ "gguf", "qwen3_5", "unsloth", "qwen", "qwen3.5", "reasoning", "chain-of-thought", "Dense", "image-text-to-text", "en", "zh", "dataset:nohurry/Opus-4.6-Reasoning-3000x-filtered", "dataset:Jackrong/Qwen3.5-reasoning-700x", "base_model:Qwen/Qwen3.5-27B", "base_model:quantized:Qwen/Qwen3.5-27B", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense multimodal transformer", "detail": "qwen3_5_text (Q4_K_M GGUF)", "total_params_b": 26.895998464, "active_params_b": 25.624600064, "layers": 64, "attention_heads": 24, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q4_K_M", "weight_bytes_per_param": 0.6149713307776606, "resident_weight_gb": 16.540267968, "active_weight_gb": 15.814117376, "kv_alloc_gb_per_1k_tokens": 0.065536, "kv_read_gb_per_1k_tokens": 0.065536, "kv_alloc_fixed_gb": 0.154927104, "kv_read_fixed_gb": 0.154927104, "notes": "Manually audited on 2026-07-06 from HF API commit 5825aa2c6c361e3ee4f588021192d1f919dd60d7, model card, served config, pinned Qwen base config comparison, linked GGUF file sizes, direct Q4_K_M GGUF header range read, and Transformers Qwen3.5 implementation review. The authoritative profile uses the card-referenced Q4_K_M artifact: 16.540267968 GB resident, 15.814117376 GB swept ordinary text traffic, 0.726150592 GB resident-only token embedding plus GGUF metadata/header/file overhead, and Qwen3.5 hybrid KV plus fixed DeltaNet state. The HF API gguf.totalFileSize points to Q2_K, so the selected artifact is intentionally pinned by card hardware/speed notes. Separate mmproj sidecar is not included unless explicitly loaded." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-GGUF/raw/5825aa2c6c361e3ee4f588021192d1f919dd60d7/README.md" }, { "label": "Model config", "url": "https://huggingface.co/Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-GGUF/raw/5825aa2c6c361e3ee4f588021192d1f919dd60d7/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-27B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-GGUF" }, { "label": "Selected Q4_K_M GGUF header audit", "url": "https://huggingface.co/Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-GGUF/resolve/5825aa2c6c361e3ee4f588021192d1f919dd60d7/Qwen3.5-27B.Q4_K_M.gguf" } ] }, { "id": "Qwen/Qwen2.5-3B-Instruct-GGUF", "name": "Qwen2.5-3B-Instruct-GGUF GGUF", "short_name": "Qwen2.5 3B Instruct", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 215418, "tags": [ "gguf", "chat", "text-generation", "en", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-3B-Instruct", "base_model:quantized:Qwen/Qwen2.5-3B-Instruct", "license:other", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 3.397103616, "active_params_b": 3.085938688, "layers": 36, "attention_heads": 16, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q5_K_M", "weight_bytes_per_param": 0.7178881363858877, "resident_weight_gb": 2.438740384, "active_weight_gb": 2.218858496, "kv_alloc_gb_per_1k_tokens": 0.036864, "kv_read_gb_per_1k_tokens": 0.036864, "notes": "Manually audited on 2026-07-06 from HF API commit 7dabda4d13d513e3e842b20f0d435c732f172cbe, model card, existing BF16 base profile, linked GGUF file sizes, and direct Q5_K_M GGUF header range read. The authoritative profile uses the card-example qwen2.5-3b-instruct-q5_k_m.gguf artifact: 2.438740384 GB resident, 2.218858496 GB swept ordinary text traffic, 0.219881888 GB resident-only token embedding plus GGUF metadata/header/file overhead, and full-context FP16 KV. The HF API gguf.totalFileSize points to q2_k, so the selected artifact is intentionally pinned by the model card download example." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-3B-Instruct-GGUF/raw/7dabda4d13d513e3e842b20f0d435c732f172cbe/README.md" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen2.5-3B-Instruct/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-3B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-3B-Instruct-GGUF" }, { "label": "Selected Q5_K_M GGUF header audit", "url": "https://huggingface.co/Qwen/Qwen2.5-3B-Instruct-GGUF/resolve/7dabda4d13d513e3e842b20f0d435c732f172cbe/qwen2.5-3b-instruct-q5_k_m.gguf" } ] }, { "id": "cyankiwi/gemma-4-E4B-it-AWQ-INT4", "name": "gemma-4-E4B-it-AWQ-INT4 4-bit", "short_name": "gemma 4 E4B it AWQ INT4", "license": "apache-2.0", "hf_pipeline_tag": "any-to-any", "hf_downloads": 247056, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "any-to-any", "base_model:google/gemma-4-E4B-it", "base_model:quantized:google/gemma-4-E4B-it", "license:apache-2.0", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 8.185960272, "active_params_b": 4.8892996, "layers": 42, "attention_heads": 8, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "compressed-tensors AWQ INT4 + F16 side tensors", "weight_bytes_per_param": 1.2596772871315982, "resident_weight_gb": 10.311668228, "active_weight_gb": 3.718346884, "kv_alloc_gb_per_1k_tokens": 0.028672, "kv_read_gb_per_1k_tokens": 0.028672, "kv_alloc_fixed_gb": 0.03670016, "kv_read_fixed_gb": 0.03670016, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, model card, served compressed-tensors config, base Google config comparison, linked-object metadata, and direct safetensors header range read. Bound weights separate the exact 10.311668228 GB resident AWQ/F16 package from the 3.718346884 GB ordinary text-decode swept subset; no lm_head tensor is stored, so the tied embedding/output projection is charged once. Hybrid Gemma 4 E4B KV uses full-context plus fixed sliding-window cache terms. The audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/cyankiwi/gemma-4-E4B-it-AWQ-INT4" }, { "label": "Model config", "url": "https://huggingface.co/cyankiwi/gemma-4-E4B-it-AWQ-INT4/raw/fc24b63a5f3313720e2d0adcdb915d7f3dc91392/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-E4B-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/cyankiwi/gemma-4-E4B-it-AWQ-INT4" }, { "label": "Pinned Google Gemma 4 E4B IT base config", "url": "https://huggingface.co/google/gemma-4-E4B-it/raw/fee6332c1abaafb77f6f9624236c63aa2f1d0187/config.json" }, { "label": "Safetensors linked-object and header audit", "url": "https://huggingface.co/cyankiwi/gemma-4-E4B-it-AWQ-INT4/resolve/fc24b63a5f3313720e2d0adcdb915d7f3dc91392/model.safetensors" } ] }, { "id": "deepseek-ai/DeepSeek-V3.2-Exp", "name": "DeepSeek-V3.2-Exp FP8", "short_name": "DeepSeek V3.2 Exp", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 229989, "tags": [ "transformers", "safetensors", "deepseek_v32", "text-generation", "conversational", "base_model:deepseek-ai/DeepSeek-V3.2-Exp-Base", "base_model:finetune:deepseek-ai/DeepSeek-V3.2-Exp-Base", "license:mit", "eval-results", "endpoints_compatible", "fp8", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "deepseek_v32 (DeepseekV32ForCausalLM)", "total_params_b": 685.396921376, "active_params_b": 40.403603136, "layers": 61, "attention_heads": 128, "kv_heads": 128, "head_dim": 56, "max_context_tokens": 163840, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "FP8 + BF16/F32 side tensors", "weight_bytes_per_param": 1.005944272139158, "resident_weight_gb": 689.4711072, "active_weight_gb": 40.403603136, "kv_alloc_gb_per_1k_tokens": 0.048068, "kv_read_gb_per_1k_tokens": 0.048068, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, model card, served config, base config, official inference code/config, safetensors index, and direct range-read headers across all 163 shards. Bound weights separate the 689.471107200 GB resident FP8/BF16/F32 package from ordinary text-decode expected-distinct MoE traffic; the next-token-prediction layer is resident-only for ordinary decode. Legacy KV fields use the conservative official DSA/MLA compressed-state coefficient. The audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp" }, { "label": "Model config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp/raw/194c67e12b1b0d6df0ef373ddcf215bc84027409/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp-Base" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/deepseek-ai/DeepSeek-V3.2-Exp" }, { "label": "Pinned base config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp-Base/raw/b21ef155036dfbbd6d49013a345d31d5f5e38c89/config.json" }, { "label": "Official inference config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp/raw/194c67e12b1b0d6df0ef373ddcf215bc84027409/inference/config_671B_v3.2.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp/raw/194c67e12b1b0d6df0ef373ddcf215bc84027409/model.safetensors.index.json" } ] }, { "id": "lmstudio-community/GLM-4.7-Flash-MLX-6bit", "name": "GLM-4.7-Flash-MLX-6bit 6-bit", "short_name": "GLM 4.7 Flash MLX 6bit", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 238894, "tags": [ "transformers", "safetensors", "glm4_moe_lite", "text-generation", "mlx", "conversational", "en", "zh", "base_model:zai-org/GLM-4.7-Flash", "base_model:quantized:zai-org/GLM-4.7-Flash", "license:mit", "endpoints_compatible", "6-bit", "region:us" ], "architecture": { "type": "Latent MLA MoE transformer", "detail": "glm4_moe_lite (Glm4MoeLiteForCausalLM)", "total_params_b": 29.94339392, "active_params_b": 3.563273871, "layers": 47, "attention_heads": 20, "kv_heads": 20, "head_dim": 256, "max_context_tokens": 202752, "routed_experts": 64, "routed_experts_per_token": 4, "shared_experts_per_token": 1, "kv_lora_rank": 512, "qk_rope_head_dim": 64, "qk_nope_head_dim": 192, "v_head_dim": 256 }, "adapter": { "kind": "moe", "weight_precision": "MLX 6-bit affine", "weight_bytes_per_param": 0.8127495265573422, "resident_weight_gb": 24.336479232, "active_weight_gb": 2.896049152, "kv_alloc_gb_per_1k_tokens": 0.054144, "kv_read_gb_per_1k_tokens": 0.054144, "expert_param_b": 0.4339771854362499, "shared_expert_param_b": 0.4339771854362499, "fixed_param_b": 1.827365129686378, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, pinned served MLX config, pinned BF16 GLM source config comparison, audited 8-bit MLX sibling comparison, safetensors index, direct shard header reads, and MLX-LM glm4_moe_lite runtime review. Header bytes record 24.336479232 GB resident, 24.058934272 GB ordinary text resident, 22.573744128 GB packed routed expert tensors, 1.485190144 GB fixed ordinary text traffic, and 0.27754496 GB resident-only embedding tensors. MLX-LM stores latent MLA cache state at 0.054144 GB per 1k context tokens; the audited profile supersedes this legacy adapter summary for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/GLM-4.7-Flash-MLX-6bit" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/GLM-4.7-Flash-MLX-6bit" }, { "label": "Pinned model config", "url": "https://huggingface.co/lmstudio-community/GLM-4.7-Flash-MLX-6bit/raw/28387239288ecde2acd4096526d6a67b3ff388f2/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/zai-org/GLM-4.7-Flash" }, { "label": "Pinned base config", "url": "https://huggingface.co/zai-org/GLM-4.7-Flash/raw/7dd20894a642a0aa287e9827cb1a1f7f91386b67/config.json" }, { "label": "Safetensors shard header audit", "url": "https://huggingface.co/lmstudio-community/GLM-4.7-Flash-MLX-6bit/raw/28387239288ecde2acd4096526d6a67b3ff388f2/model.safetensors.index.json" }, { "label": "MLX-LM GLM4 MoE Lite implementation", "url": "https://github.com/ml-explore/mlx-lm/blob/2ed22318cd6a2fcc5c2e0caa1e1fb0ddeb7cafd5/mlx_lm/models/glm4_moe_lite.py" } ] }, { "id": "Qwen/Qwen2.5-Coder-1.5B", "name": "Qwen2.5-Coder-1.5B BF16", "short_name": "Qwen2.5 Coder 1.5B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 189380, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "code", "qwen", "qwen-coder", "codeqwen", "conversational", "en", "arxiv:2409.12186", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-1.5B", "base_model:finetune:Qwen/Qwen2.5-1.5B", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 1.543714304, "active_params_b": 1.543714304, "layers": 28, "attention_heads": 12, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 3.087428608, "active_weight_gb": 3.087428608, "kv_alloc_gb_per_1k_tokens": 0.028672, "kv_read_gb_per_1k_tokens": 0.028672, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, pinned served Coder config, Qwen2.5 base config comparison, audited Coder Instruct config comparison, and a direct safetensors header range read. The header records 338 BF16 tensors, 1.543714304B parameters, 3.087428608 GB resident and swept BF16 decode traffic, no stored lm_head.weight, and a tied embedding/output table. Full-context BF16 KV traffic is 0.028672 GB per 1k context tokens." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B" }, { "label": "Pinned model config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B/raw/df3ce67c0e24480f20468b6ef2894622d69eb73b/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Coder-1.5B" }, { "label": "Pinned base config", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B/raw/8faed761d45a263340a0528343f099c05c9a4323/config.json" }, { "label": "Pinned Coder Instruct config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B-Instruct/raw/2e1fd397ee46e1388853d2af2c993145b0f1098a/config.json" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B/resolve/df3ce67c0e24480f20468b6ef2894622d69eb73b/model.safetensors" } ] }, { "id": "bartowski/Llama-3.2-3B-Instruct-GGUF", "name": "Llama-3.2-3B-Instruct-GGUF GGUF", "short_name": "Llama 3.2 3B Instruct", "license": "llama3.2", "hf_pipeline_tag": "text-generation", "hf_downloads": 213102, "tags": [ "gguf", "facebook", "meta", "llama", "llama-3", "text-generation", "en", "de", "fr", "it", "pt", "hi", "es", "th", "base_model:meta-llama/Llama-3.2-3B-Instruct", "base_model:quantized:meta-llama/Llama-3.2-3B-Instruct", "license:llama3.2", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "llama GGUF F16 (Llama 3.2 3B Instruct)", "total_params_b": 3.212749888, "active_params_b": 3.212749888, "layers": 28, "attention_heads": 24, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF F16", "weight_bytes_per_param": 2, "resident_weight_gb": 6.43368784, "active_weight_gb": 6.425850112, "kv_alloc_gb_per_1k_tokens": 0.114688, "kv_read_gb_per_1k_tokens": 0.114688, "notes": "Manually audited on 2026-07-06 from current HF API metadata, model card, base-model API metadata, gated base-config access check, linked-object HEAD checks, and a direct selected-file GGUF header range read. The authoritative bounds profile uses the API-selected Llama-3.2-3B-Instruct-f16.gguf artifact: 6.433687840 GB resident, 6.425850112 GB swept tensor traffic, 0.007837728 GB resident-only GGUF metadata/header/file overhead, 28 layers, 8 KV heads, 128 key/value head dimension, and 131072-token context." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/bartowski/Llama-3.2-3B-Instruct-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/bartowski/Llama-3.2-3B-Instruct-GGUF" }, { "label": "Base model card", "url": "https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct" }, { "label": "Base model API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.2-3B-Instruct" }, { "label": "Selected F16 GGUF header audit", "url": "https://huggingface.co/bartowski/Llama-3.2-3B-Instruct-GGUF/resolve/5ab33fa94d1d04e903623ae72c95d1696f09f9e8/Llama-3.2-3B-Instruct-f16.gguf" } ] }, { "id": "lmstudio-community/gemma-4-26B-A4B-it-MLX-8bit", "name": "gemma-4-26B-A4B-it-MLX-8bit 8-bit", "short_name": "gemma 4 26B A4B it MLX 8bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 208960, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "mlx", "conversational", "base_model:google/gemma-4-26B-A4B-it", "base_model:quantized:google/gemma-4-26B-A4B-it", "license:apache-2.0", "endpoints_compatible", "8-bit", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "gemma4_text", "total_params_b": 26.544131376, "active_params_b": 4, "layers": 30, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "MLX 8-bit affine", "weight_bytes_per_param": 1.0530921873478298, "resident_weight_gb": 27.953417372, "active_weight_gb": 4.06209542, "kv_alloc_gb_per_1k_tokens": 0.20992, "kv_read_gb_per_1k_tokens": 0.20992, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, served MLX config, non-QAT Google config comparison, audited MLX 4-bit sibling config comparison, safetensors index, and direct range-read shard headers. Bound weights separate the exact 27.953417372 GB resident MLX package from 2.545510460 GB fixed ordinary text traffic plus 0.189573120 GB per routed expert index; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-MLX-8bit" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-26B-A4B-it-MLX-8bit" }, { "label": "Pinned MLX model config", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-MLX-8bit/raw/669237cd8dad363224c976432475b81dd5db5a89/config.json" }, { "label": "Pinned Google base config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it/raw/20da991ab4afab98e8f910c4a2e8f4fbefc404ad/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-MLX-8bit/raw/669237cd8dad363224c976432475b81dd5db5a89/model.safetensors.index.json" } ] }, { "id": "google/paligemma-3b-mix-224", "name": "paligemma-3b-mix-224 F32", "short_name": "paligemma 3b mix 224", "license": "gemma", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 227287, "tags": [ "transformers", "safetensors", "paligemma", "image-text-to-text", "arxiv:2310.09199", "arxiv:2303.15343", "arxiv:2403.08295", "arxiv:1706.03762", "arxiv:2010.11929", "arxiv:2209.06794", "arxiv:2209.04372", "arxiv:2103.01913", "arxiv:2205.12522", "arxiv:2110.11624", "arxiv:2108.03353", "arxiv:2010.04295", "arxiv:2401.06209", "arxiv:2305.10355", "arxiv:2203.10244", "arxiv:1810.12440", "arxiv:1905.13648", "arxiv:1608.00272", "arxiv:1908.04913", "arxiv:2407.07726", "license:gemma", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Unknown gated multimodal model", "detail": "PaliGemmaForConditionalGeneration (gated config unavailable)", "total_params_b": 2.92346648, "active_params_b": 2.92346648, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "F32", "weight_bytes_per_param": 4, "resident_weight_gb": 11.69386592, "active_weight_gb": 11.69386592, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Fail-closed on 2026-07-06 from current Hugging Face API metadata and gated-access checks. The API records 2923466480 F32 safetensors parameters at commit d1d8734c9c3ad0ccfeea4afc270faa356c2ba515, but raw README, config, model.safetensors.index.json, ranged safetensors shard bytes, and hf download config.json all require manual approval in this environment. KV geometry, exact context, text/vision tensor split, tied embeddings, and swept decode traffic are intentionally unknown until direct gated evidence is available." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/paligemma-3b-mix-224" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/paligemma-3b-mix-224" }, { "label": "Pinned gated config access check", "url": "https://huggingface.co/google/paligemma-3b-mix-224/raw/d1d8734c9c3ad0ccfeea4afc270faa356c2ba515/config.json" } ] }, { "id": "microsoft/DialoGPT-medium", "name": "DialoGPT-medium FP16", "short_name": "DialoGPT medium", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 233570, "tags": [ "transformers", "pytorch", "tf", "jax", "rust", "gpt2", "text-generation", "conversational", "arxiv:1911.00536", "license:mit", "text-generation-inference", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gpt2 (GPT2LMHeadModel)", "total_params_b": 0.354823168, "active_params_b": 0.353774592, "layers": 24, "attention_heads": 16, "kv_heads": 16, "head_dim": 64, "max_context_tokens": 1024, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "FP16", "weight_bytes_per_param": 2, "resident_weight_gb": 0.709646336, "active_weight_gb": 0.707549184, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, pinned config, model card, Transformers 4.57.1 GPT-2 implementation, pinned PyTorch checkpoint tensor grouping, and actual AutoModelForCausalLM load-time weight tying behavior. The default PyTorch checkpoint loads as FP16 with 0.709646336 GB unique serving parameters; ordinary decode sweeps 0.707549184 GB after lm_head is tied to transformer.wte, while transformer.wpe.weight is resident-only." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/microsoft/DialoGPT-medium" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/microsoft/DialoGPT-medium" }, { "label": "Pinned model config", "url": "https://huggingface.co/microsoft/DialoGPT-medium/raw/7b40bb0f92c45fefa957d088000d8648e5c7fa33/config.json" }, { "label": "Pinned PyTorch checkpoint audit", "url": "https://huggingface.co/microsoft/DialoGPT-medium/resolve/7b40bb0f92c45fefa957d088000d8648e5c7fa33/pytorch_model.bin" } ] }, { "id": "Qwen/Qwen3.5-2B-Base", "name": "Qwen3.5-2B-Base F32/BF16", "short_name": "Qwen3.5 2B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 231586, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "conversational", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 2.274069824, "active_params_b": 1.881825088, "layers": 24, "attention_heads": 8, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "F32/BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 4.548144832, "active_weight_gb": 3.76365536, "kv_alloc_gb_per_1k_tokens": 0.049152, "kv_read_gb_per_1k_tokens": 0.049152, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, pinned config, Qwen/Qwen3.5-2B config comparison, safetensors index, direct safetensors header range read, and the Transformers qwen3_5 implementation. Bound weights separate the 4.548144832 GB resident BF16/F32 package from the 3.76365536 GB ordinary text-decode swept subset; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3.5-2B-Base" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-2B-Base" }, { "label": "Pinned model config", "url": "https://huggingface.co/Qwen/Qwen3.5-2B-Base/raw/b1485b2fa6dfa1287294f269f5fb618e03d52d7c/config.json" }, { "label": "Safetensors index", "url": "https://huggingface.co/Qwen/Qwen3.5-2B-Base/raw/b1485b2fa6dfa1287294f269f5fb618e03d52d7c/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen3-30B-A3B-Thinking-2507", "name": "Qwen3-30B-A3B-Thinking-2507 BF16", "short_name": "Qwen3 30B A3B Thinking 2507", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 212751, "tags": [ "transformers", "safetensors", "qwen3_moe", "text-generation", "conversational", "arxiv:2402.17463", "arxiv:2407.02490", "arxiv:2501.15383", "arxiv:2404.06654", "arxiv:2505.09388", "license:apache-2.0", "eval-results", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_moe (Qwen3MoeForCausalLM)", "total_params_b": 30.532122624, "active_params_b": 3.3, "layers": 48, "attention_heads": 32, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 61.064245248, "active_weight_gb": 6.083735552, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, model card, pinned served config, sibling config comparison, safetensors index, and direct shard header range reads. Bound weights separate the exact 61.064245248 GB resident BF16 package from the 6.083735552 GB ordinary single-session MoE text-decode traffic; the audited profile supersedes the generated rounded active-weight estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Thinking-2507" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Thinking-2507/raw/144afc2f379b542fdd4e85a1fcd5e1f79112d95d/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-30B-A3B-Thinking-2507" }, { "label": "Safetensors index", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-Thinking-2507/raw/144afc2f379b542fdd4e85a1fcd5e1f79112d95d/model.safetensors.index.json" } ] }, { "id": "allenai/Olmo-3-7B-Instruct-SFT", "name": "Olmo-3-7B-Instruct-SFT BF16", "short_name": "Olmo 3 7B Instruct SFT", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 214141, "tags": [ "transformers", "safetensors", "olmo3", "text-generation", "conversational", "en", "dataset:allenai/Dolci-Instruct-SFT", "dataset:allenai/Dolci-Instruct-SFT-Tool-Use-SA", "arxiv:2512.13961", "base_model:allenai/Olmo-3-1025-7B", "base_model:finetune:allenai/Olmo-3-1025-7B", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "olmo3 (Olmo3ForCausalLM)", "total_params_b": 7.298011136, "active_params_b": 7.298011136, "layers": 32, "attention_heads": 32, "kv_heads": 32, "head_dim": 128, "max_context_tokens": 65536, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 14.596022272, "active_weight_gb": 13.774544896, "kv_alloc_gb_per_1k_tokens": 0.524288, "kv_read_gb_per_1k_tokens": 0.524288, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, pinned model card, pinned config, base config comparison, Transformers OLMo3 implementation, linked object metadata, safetensors index, and direct safetensors shard-header range reads. Direct headers found 14.596022272 GB resident BF16 tensor payload, 13.774544896 GB ordinary text-decode swept traffic, and 0.821477376 GB resident-only input embedding. The authoritative profile uses layered KV: 8 full-context layers plus 24 sliding-window layers capped at 4096 tokens." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/allenai/Olmo-3-7B-Instruct-SFT" }, { "label": "Model config", "url": "https://huggingface.co/allenai/Olmo-3-7B-Instruct-SFT/raw/e1452fc572d51966ff4aaeb25118b891eb93e549/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/allenai/Olmo-3-7B-Instruct-SFT" }, { "label": "Base model card", "url": "https://huggingface.co/allenai/Olmo-3-1025-7B" }, { "label": "Safetensors index and shard-header audit", "url": "https://huggingface.co/allenai/Olmo-3-7B-Instruct-SFT/raw/e1452fc572d51966ff4aaeb25118b891eb93e549/model.safetensors.index.json" } ] }, { "id": "pytorch/gemma-3-27b-it-AWQ-INT4", "name": "gemma-3-27b-it-AWQ-INT4 4-bit", "short_name": "gemma 3 27b it AWQ INT4", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 184232, "tags": [ "transformers", "pytorch", "gemma3", "image-text-to-text", "torchao", "conversational", "en", "arxiv:2507.16099", "base_model:google/gemma-3-27b-it", "base_model:quantized:google/gemma-3-27b-it", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma3_text (TorchAO AWQ INT4)", "total_params_b": 27.43240664, "active_params_b": 27.009346304, "layers": 62, "attention_heads": 32, "kv_heads": 16, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "TorchAO AWQ INT4", "weight_bytes_per_param": 0.6296832584426869, "resident_weight_gb": 17.2737272, "active_weight_gb": 16.427606528, "kv_alloc_gb_per_1k_tokens": 0.08628207616, "kv_read_gb_per_1k_tokens": 0.095551488, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, model card, pinned served config, PyTorch index metadata, linked-object HEAD checks, byte-range ZIP64 central-directory reads, and direct data.pkl metadata parsing. downloads=184232. Bound weights separate the exact 17.273727200 GB PyTorch archive storage from 16.427606528 GB ordinary text-decode swept traffic. The tied language_model.model.embed_tokens.weight and language_model.lm_head.weight entries share one 2.819260416 GB BF16 storage blob, charged once in resident memory and once in swept output-projection traffic. Hybrid Gemma 3 KV uses full-context plus fixed sliding-window BF16 cache terms. The audited profile supersedes the generated ideal 0.5-byte estimate." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/pytorch/gemma-3-27b-it-AWQ-INT4" }, { "label": "Pinned model config", "url": "https://huggingface.co/pytorch/gemma-3-27b-it-AWQ-INT4/raw/0e6f8915f5b53333c13bbd48fc6ead75cc7625d7/config.json" }, { "label": "PyTorch index and archive metadata audit", "url": "https://huggingface.co/pytorch/gemma-3-27b-it-AWQ-INT4/raw/0e6f8915f5b53333c13bbd48fc6ead75cc7625d7/pytorch_model.bin.index.json" } ] }, { "id": "QuantTrio/Qwen3-VL-32B-Instruct-AWQ", "name": "Qwen3-VL-32B-Instruct-AWQ 4-bit", "short_name": "Qwen3 VL 32B Instruct", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 259994, "tags": [ "transformers", "safetensors", "qwen3_vl", "image-text-to-text", "AWQ", "vLLM", "conversational", "arxiv:2505.09388", "arxiv:2502.13923", "arxiv:2409.12191", "arxiv:2308.12966", "base_model:Qwen/Qwen3-VL-32B-Instruct", "base_model:quantized:Qwen/Qwen3-VL-32B-Instruct", "license:apache-2.0", "endpoints_compatible", "4-bit", "awq", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_vl_text", "total_params_b": 33.357390064, "active_params_b": 31.984210944, "layers": 64, "attention_heads": 64, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "AWQ INT4", "weight_bytes_per_param": 0.6150310995146205, "resident_weight_gb": 20.515832288, "active_weight_gb": 17.769474048, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, model card, pinned served AWQ config, audited BF16 base profile/config, official FP8 sibling config comparison, linked-object HEAD checks, safetensors index, and direct shard-header range reads. downloads=259994. Bound weights separate the exact 20.515832288 GB AWQ/F16 package from 17.769474048 GB ordinary text-decode swept traffic; model.visual plus input embeddings remain resident-only. The audited profile supersedes the generated ideal 0.5-byte estimate." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/QuantTrio/Qwen3-VL-32B-Instruct-AWQ" }, { "label": "Pinned model config", "url": "https://huggingface.co/QuantTrio/Qwen3-VL-32B-Instruct-AWQ/raw/c2d44b1376c73fa120fdb4e216c2ef61c1945498/config.json" }, { "label": "Safetensors index and shard-header audit", "url": "https://huggingface.co/QuantTrio/Qwen3-VL-32B-Instruct-AWQ/raw/c2d44b1376c73fa120fdb4e216c2ef61c1945498/model.safetensors.index.json" } ] }, { "id": "XiaomiMiMo/MiMo-V2.5", "name": "MiMo-V2.5 FP8", "short_name": "MiMo V2.5", "license": "mit", "hf_pipeline_tag": "unknown", "hf_downloads": 219382, "tags": [ "safetensors", "mimo_v2", "multimodal", "vision-language", "audio", "agent", "video-understanding", "long-context", "custom_code", "en", "zh", "license:mit", "eval-results", "fp8", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "mimo_v2 (MiMoV2ForCausalLM)", "total_params_b": 310.77504, "active_params_b": 14.82175072, "layers": 48, "attention_heads": 64, "kv_heads": 8, "global_kv_heads": 4, "head_dim": 192, "max_context_tokens": 1048576, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "FP8", "weight_bytes_per_param": 1.013694993677742, "resident_weight_gb": 315.031102208, "active_weight_gb": 17.20789696, "kv_alloc_gb_per_1k_tokens": 0.0232955904, "kv_read_gb_per_1k_tokens": 0.02383872, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, pinned model card, pinned served config, custom modeling code review, safetensors index, and direct shard-header range reads. downloads=219382. Bound weights separate the exact 315.031102208 GB FP8/BF16/F32 package from 17.207896960 GB ordinary single-session text-decode active traffic. Resident-only input embedding, visual, audio, and MTP tensors total 4.418746240 GB. KV accounting uses the runnable hybrid full-context plus 128-token sliding-window layout instead of the generated all-full-context estimate." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/XiaomiMiMo/MiMo-V2.5" }, { "label": "Pinned model config", "url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.5/raw/2fd4f899a491de2fb0beeafe32b5d700b251f593/config.json" }, { "label": "Safetensors index and shard-header audit", "url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.5/raw/2fd4f899a491de2fb0beeafe32b5d700b251f593/model.safetensors.index.json" } ] }, { "id": "Jackrong/Qwopus3.5-9B-Coder-MTP-GGUF", "name": "Qwopus3.5-9B-Coder-MTP-GGUF GGUF", "short_name": "Qwopus3.5 9B Coder MTP", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 139234, "tags": [ "transformers", "gguf", "llama.cpp", "image-text-to-text", "vision", "multimodal", "text-generation-inference", "unsloth", "conversational", "mtp", "multi-token-prediction", "speculative-decoding", "qwen3_5", "reasoning", "chain-of-thought", "lora", "sft", "agent", "coder", "tool-use", "function-calling", "en", "zh", "ko", "ru", "ja", "es", "dataset:Jackrong/Claude-opus-4.7-TraceInversion-5000x", "dataset:Jackrong/Claude-opus-4.6-TraceInversion-9000x", "base_model:Jackrong/Qwopus3.5-9B-v3.5", "base_model:adapter:Jackrong/Qwopus3.5-9B-v3.5", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 9.197093888, "active_params_b": 7.936684544, "layers": 32, "attention_heads": 16, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q5_K_M", "weight_bytes_per_param": 0.7222440776283615, "resident_weight_gb": 6.642546592, "active_weight_gb": 5.757732864, "kv_alloc_gb_per_1k_tokens": 0.032768, "kv_read_gb_per_1k_tokens": 0.032768, "kv_alloc_fixed_gb": 0.051904512, "kv_read_fixed_gb": 0.051904512, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, model card, Jackrong base config, Qwen3.5 base config comparison, selected linked-object HEAD checks, direct GGUF header/tensor-index range read, and Qwen3.5 runtime review. downloads=139234. The API GGUF block points at mmproj-F32, so the audited profile uses the selected Q5_K_M main artifact and supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Jackrong/Qwopus3.5-9B-Coder-MTP-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Jackrong/Qwopus3.5-9B-Coder-MTP-GGUF" }, { "label": "Selected Q5_K_M GGUF artifact", "url": "https://huggingface.co/Jackrong/Qwopus3.5-9B-Coder-MTP-GGUF/resolve/bf3fdbf18a25cf1ce449589b23c067ca29d6d1a9/Qwopus3.5-9B-Coder-MTP-Q5_K_M.gguf" }, { "label": "Base model config", "url": "https://huggingface.co/Jackrong/Qwopus3.5-9B-v3.5/raw/dc2b00e1b1bc404133e3a3e15e7ddcdff814fd86/config.json" }, { "label": "Qwen3.5 base config", "url": "https://huggingface.co/Qwen/Qwen3.5-9B/raw/c202236235762e1c871ad0ccb60c8ee5ba337b9a/config.json" }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py" } ] }, { "id": "Qwen/Qwen3-235B-A22B-Instruct-2507-FP8", "name": "Qwen3-235B-A22B-Instruct-2507-FP8 FP8", "short_name": "Qwen3 235B A22B Instruct 2507 FP8", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 221064, "tags": [ "transformers", "safetensors", "qwen3_moe", "text-generation", "conversational", "arxiv:2505.09388", "base_model:Qwen/Qwen3-235B-A22B-Instruct-2507", "base_model:quantized:Qwen/Qwen3-235B-A22B-Instruct-2507", "license:apache-2.0", "endpoints_compatible", "fp8", "deploy:azure", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_moe (Qwen3MoeForCausalLM)", "total_params_b": 235.107904512, "active_params_b": 21.569709056, "layers": 94, "attention_heads": 64, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "FP8/BF16", "weight_bytes_per_param": 1.0055677028244416, "resident_weight_gb": 236.416915456, "active_weight_gb": 22.243395584, "kv_alloc_gb_per_1k_tokens": 0.192512, "kv_read_gb_per_1k_tokens": 0.192512, "notes": "Manual correction on 2026-07-06 from live Hugging Face API metadata, pinned model card, served FP8 config, BF16 base config comparison, safetensors index, and direct safetensors header range reads for all 24 shards. The authoritative bounds profile uses exact stored bytes: 236.416915456 GB resident, 8.048138240 GB fixed ordinary text traffic, 1.774407168 GB per routed expert, 8 routed experts per token, 128 routed experts, no shared expert, and full-context BF16 KV." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507-FP8" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507-FP8/raw/e156cb4efae43fbee1a1ab073f946a1377e6b969/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-235B-A22B-Instruct-2507-FP8" }, { "label": "Pinned safetensors index and header audit", "url": "https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507-FP8/raw/e156cb4efae43fbee1a1ab073f946a1377e6b969/model.safetensors.index.json" } ] }, { "id": "unsloth/gemma-4-E4B-it-qat-GGUF", "name": "gemma-4-E4B-it-qat-GGUF GGUF 4-bit", "short_name": "gemma 4 E4B it qat", "license": "apache-2.0", "hf_pipeline_tag": "any-to-any", "hf_downloads": 236805, "tags": [ "transformers", "gguf", "gemma4", "image-text-to-text", "unsloth", "gemma", "google", "any-to-any", "base_model:google/gemma-4-E4B-it-qat-q4_0-unquantized", "base_model:quantized:google/gemma-4-E4B-it-qat-q4_0-unquantized", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 7.463013674, "active_params_b": 4.644441386, "layers": 42, "attention_heads": 8, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF UD-Q4_K_XL", "weight_bytes_per_param": 0.5648782039200643, "resident_weight_gb": 4.21569376, "active_weight_gb": 2.61442592, "kv_alloc_gb_per_1k_tokens": 0.028672, "kv_read_gb_per_1k_tokens": 0.028672, "kv_alloc_fixed_gb": 0.03670016, "kv_read_fixed_gb": 0.03670016, "notes": "Manual correction on 2026-07-06 from live Hugging Face API metadata, pinned model card, pinned package config, pinned Google QAT unquantized base config, linked-object HEAD checks, and direct GGUF header range read of the selected UD-Q4_K_XL artifact. The API gguf.totalFileSize currently matches the smaller UD-Q2_K_XL sibling, so this row follows the model-card-selected UD-Q4_K_XL artifact. Bound weights use the selected file size 4.215693760 GB for residency and 2.614425920 GB of ordinary swept tensor traffic after excluding the resident per-layer token embedding lookup table." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/gemma-4-E4B-it-qat-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/unsloth/gemma-4-E4B-it-qat-GGUF/raw/bbcd9d849c2541ecc2af7ef64b3c3c2c7aa14e96/config.json" }, { "label": "Base model config", "url": "https://huggingface.co/google/gemma-4-E4B-it-qat-q4_0-unquantized/raw/dfc5b925ddb1d41aaf1fe9679abdcfb0805e1aa6/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/gemma-4-E4B-it-qat-GGUF" }, { "label": "Selected GGUF header audit", "url": "https://huggingface.co/unsloth/gemma-4-E4B-it-qat-GGUF/resolve/bbcd9d849c2541ecc2af7ef64b3c3c2c7aa14e96/gemma-4-E4B-it-qat-UD-Q4_K_XL.gguf" } ] }, { "id": "AEON-7/Qwen3.6-35B-A3B-heretic-NVFP4", "name": "Qwen3.6-35B-A3B-heretic-NVFP4 NVFP4", "short_name": "Qwen3.6 35B A3B heretic NVFP4", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 203785, "tags": [ "transformers", "safetensors", "qwen3_5_moe", "image-text-to-text", "1m-context", "35b", "3b-active", "4-bit", "aarch64", "abliterated", "aeon", "aeon-7", "agentic", "arm64", "blackwell", "chat", "chunked-prefill", "coding", "compressed-tensors", "conversational", "decensored", "dflash", "dgx-spark", "english", "flashinfer-cutlass", "fp4", "fp8-kv-cache", "function-calling", "gated-deltanet", "gb10", "gdn", "gpu", "grace-blackwell", "heretic", "hybrid", "hybrid-attention", "instruct", "linear-attention", "llmcompressor", "long-context", "mixture-of-experts", "moe", "multimodal", "native-fp4", "nvfp4", "openai-api", "openai-compatible", "prefix-caching", "production-ready", "quantized", "qwen", "qwen3", "qwen3.5", "qwen3.6", "reasoning", "refusal-removed", "sm_121a", "sparse", "sparse-moe", "speculative-decoding", "thinking", "tool-calling", "uncensored", "unfiltered", "vision", "vision-language", "vllm", "weight-quantization", "en", "base_model:tvall43/Qwen3.6-35B-A3B-heretic", "base_model:quantized:tvall43/Qwen3.6-35B-A3B-heretic", "license:apache-2.0", "endpoints_compatible", "8-bit", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_5_moe_text", "total_params_b": 35.1072, "active_params_b": 3, "layers": 40, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "NVFP4", "weight_bytes_per_param": 0.5, "resident_weight_gb": 23.336710112, "active_weight_gb": 3.873048576, "kv_alloc_gb_per_1k_tokens": 0.02048, "kv_read_gb_per_1k_tokens": 0.02048, "kv_alloc_fixed_gb": 0.06488064, "kv_read_fixed_gb": 0.06488064, "notes": "Audited from Hugging Face metadata and direct safetensors header reads on 2026-07-06; downloads=203785; data quality: config, params, base:tvall43/Qwen3.6-35B-A3B-heretic, exact tensor bytes. Bounds use the self-contained AEON-7 profile for exact expected-distinct MoE traffic; this catalog adapter is only a visible fallback summary." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/AEON-7/Qwen3.6-35B-A3B-heretic-NVFP4" }, { "label": "Model config", "url": "https://huggingface.co/AEON-7/Qwen3.6-35B-A3B-heretic-NVFP4/raw/3fce8ab8e76a49f2bd1701d4fbfba021ef2143dd/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/tvall43/Qwen3.6-35B-A3B-heretic" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/AEON-7/Qwen3.6-35B-A3B-heretic-NVFP4/resolve/3fce8ab8e76a49f2bd1701d4fbfba021ef2143dd/model.safetensors" } ] }, { "id": "Qwen/Qwen2.5-VL-72B-Instruct-AWQ", "name": "Qwen2.5-VL-72B-Instruct-AWQ 4-bit", "short_name": "Qwen2.5 VL 72B Instruct", "license": "other", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 184780, "tags": [ "transformers", "safetensors", "qwen2_5_vl", "image-text-to-text", "multimodal", "conversational", "en", "arxiv:2309.00071", "arxiv:2409.12191", "arxiv:2308.12966", "base_model:Qwen/Qwen2.5-VL-72B-Instruct", "base_model:quantized:Qwen/Qwen2.5-VL-72B-Instruct", "license:other", "text-generation-inference", "endpoints_compatible", "4-bit", "awq", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2_5_vl (Qwen2_5_VLForConditionalGeneration)", "total_params_b": 73.6624, "active_params_b": 73.6624, "layers": 80, "attention_heads": 64, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 128000, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "4-bit", "weight_bytes_per_param": 0.5, "resident_weight_gb": 43.004666368, "active_weight_gb": 39.1041024, "kv_alloc_gb_per_1k_tokens": 0.32768, "kv_read_gb_per_1k_tokens": 0.32768, "notes": "Audited from Hugging Face metadata and direct safetensors shard-header reads on 2026-07-06; downloads=184780; data quality: config, params, base:Qwen/Qwen2.5-VL-72B-Instruct, exact tensor bytes. Resident bytes include AWQ qweights/qzeros/scales plus BF16 visual, embedding, and head tensors; active bytes exclude the resident-only input embedding and visual tower." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-VL-72B-Instruct-AWQ" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-72B-Instruct-AWQ/raw/c8b87d4b81f34b6a147577a310d7e75f0698f6c2/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-72B-Instruct" }, { "label": "Safetensors index and shard header audit", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-72B-Instruct-AWQ/raw/c8b87d4b81f34b6a147577a310d7e75f0698f6c2/model.safetensors.index.json" } ] }, { "id": "huihui-ai/Huihui-DeepSeek-V4-Flash-abliterated-ds4-GGUF", "name": "Huihui-DeepSeek-V4-Flash-abliterated-ds4-GGUF GGUF Q2_K", "short_name": "Huihui DeepSeek V4 Flash abliterated ds4", "license": "mit", "hf_pipeline_tag": "unknown", "hf_downloads": 584606, "tags": [ "transformers", "gguf", "abliterated", "uncensored", "GGUF", "quantized", "deepseek", "deepseek-v4", "deepseek-v4-flash", "moe", "mixture-of-experts", "2-bit", "4-bit", "iq2_xxs", "q2_k", "q4_k", "ds4", "apple-silicon", "metal", "llama.cpp", "base_model:deepseek-ai/DeepSeek-V4-Flash", "base_model:quantized:deepseek-ai/DeepSeek-V4-Flash", "license:mit", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "MoE transformer", "detail": "deepseek_v4 (DeepseekV4ForCausalLM)", "total_params_b": 158.0694, "active_params_b": 3.7048, "layers": 43, "attention_heads": 64, "kv_heads": 1, "head_dim": 512, "max_context_tokens": 1048576, "routed_experts": 256, "routed_experts_per_token": 6, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "GGUF Q2_K", "weight_bytes_per_param": 0.35066322486499185, "resident_weight_gb": 99.705676384, "active_weight_gb": 9.872767324, "kv_alloc_gb_per_1k_tokens": 0.00314, "kv_read_gb_per_1k_tokens": 0.00121875, "notes": "Audited from current Hugging Face metadata and selected GGUF tensor header on 2026-07-06; downloads=584606; selected artifact=Huihui-DeepSeek-V4-Flash-BF16-abliterated-ds4-Q2_K.gguf. Resident bytes are exact linked-object bytes; active weight is exact single-session ordinary DS4 decode traffic through fixed tensors plus six routed experts; KV/state coefficients use the audited DeepSeek V4 DS4 compressed-state profile." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/huihui-ai/Huihui-DeepSeek-V4-Flash-abliterated-ds4-GGUF" }, { "label": "Hugging Face model card", "url": "https://huggingface.co/huihui-ai/Huihui-DeepSeek-V4-Flash-abliterated-ds4-GGUF/blob/f06f59bce3c36b3282b75c9fe2621c83c9399d10/README.md" }, { "label": "Base model config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash/raw/60d8d70770c6776ff598c94bb586a859a38244f1/config.json" }, { "label": "Selected GGUF tensor header audit", "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" } ] }, { "id": "lmstudio-community/gemma-4-26B-A4B-it-MLX-6bit", "name": "gemma-4-26B-A4B-it-MLX-6bit 6-bit", "short_name": "gemma 4 26B A4B it MLX 6bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 200715, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "mlx", "conversational", "base_model:google/gemma-4-26B-A4B-it", "base_model:quantized:google/gemma-4-26B-A4B-it", "license:apache-2.0", "endpoints_compatible", "6-bit", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "gemma4_text", "total_params_b": 26.544131376, "active_params_b": 4, "layers": 30, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "MLX 6-bit affine", "weight_bytes_per_param": 0.8205586161200742, "resident_weight_gb": 21.781015708, "active_weight_gb": 3.243157564, "kv_alloc_gb_per_1k_tokens": 0.20992, "kv_read_gb_per_1k_tokens": 0.20992, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, served MLX config, non-QAT Google config comparison, audited MLX sibling config comparison, safetensors index, and direct range-read shard headers. Bound weights separate the exact 21.781015708 GB resident MLX package from 2.083416124 GB fixed ordinary text traffic plus 0.144967680 GB per routed expert index; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-MLX-6bit" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-26B-A4B-it-MLX-6bit" }, { "label": "Pinned MLX model config", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-MLX-6bit/raw/fd6c729deddbd6211b37be13f2a42ef2603b3b56/config.json" }, { "label": "Pinned Google base config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it/raw/20da991ab4afab98e8f910c4a2e8f4fbefc404ad/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-MLX-6bit/raw/fd6c729deddbd6211b37be13f2a42ef2603b3b56/model.safetensors.index.json" } ] }, { "id": "lmstudio-community/gemma-4-26B-A4B-it-MLX-5bit", "name": "gemma-4-26B-A4B-it-MLX-5bit 5-bit", "short_name": "gemma 4 26B A4B it MLX 5bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 199805, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "mlx", "conversational", "base_model:google/gemma-4-26B-A4B-it", "base_model:quantized:google/gemma-4-26B-A4B-it", "license:apache-2.0", "endpoints_compatible", "5-bit", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "gemma4_text", "total_params_b": 26.544131376, "active_params_b": 4, "layers": 30, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "MLX 5-bit affine", "weight_bytes_per_param": 0.7042918305061963, "resident_weight_gb": 18.694814876, "active_weight_gb": 2.833688636, "kv_alloc_gb_per_1k_tokens": 0.20992, "kv_read_gb_per_1k_tokens": 0.20992, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, served MLX config, non-QAT Google config comparison, audited MLX sibling config comparison, safetensors index, and direct range-read shard headers. Bound weights separate the exact 18.694814876 GB resident MLX package from 1.852368956 GB fixed ordinary text traffic plus 0.122664960 GB per routed expert index; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-MLX-5bit" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-26B-A4B-it-MLX-5bit" }, { "label": "Pinned MLX model config", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-MLX-5bit/raw/13c9cdfe77d1aa897e7e94ab7d418ddf2345f219/config.json" }, { "label": "Pinned Google base config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it/raw/20da991ab4afab98e8f910c4a2e8f4fbefc404ad/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-MLX-5bit/raw/13c9cdfe77d1aa897e7e94ab7d418ddf2345f219/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen3.5-9B-Base", "name": "Qwen3.5-9B-Base BF16/F32", "short_name": "Qwen3.5 9B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 198528, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "conversational", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 9.653104368, "active_params_b": 7.936684544, "layers": 32, "attention_heads": 16, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16/F32", "weight_bytes_per_param": 2.000000795598981, "resident_weight_gb": 19.306216416, "active_weight_gb": 15.873376768, "kv_alloc_gb_per_1k_tokens": 0.032768, "kv_read_gb_per_1k_tokens": 0.032768, "kv_alloc_fixed_gb": 0.051904512, "kv_read_fixed_gb": 0.051904512, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, pinned model card/config, instruction sibling config comparison, safetensors index, direct shard header range reads, and the Transformers qwen3_5 implementation. Bound weights separate the exact 19.306216416 GB resident BF16/F32 package from the 15.873376768 GB ordinary text-decode swept traffic; the audited profile supersedes the generated full-context-only KV estimate and adds the fixed DeltaNet runtime state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3.5-9B-Base" }, { "label": "Pinned Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-9B-Base" }, { "label": "Pinned model config", "url": "https://huggingface.co/Qwen/Qwen3.5-9B-Base/raw/68c46c4b3498877f3ef123c856ecfde50c39f404/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Qwen/Qwen3.5-9B-Base/raw/68c46c4b3498877f3ef123c856ecfde50c39f404/model.safetensors.index.json" } ] }, { "id": "lmstudio-community/Qwen3-Coder-Next-MLX-8bit", "name": "Qwen3-Coder-Next-MLX-8bit 8-bit", "short_name": "Qwen3 Coder Next MLX 8bit", "license": "apache-2.0", "hf_pipeline_tag": "unknown", "hf_downloads": 203906, "tags": [ "mlx", "safetensors", "qwen3_next", "base_model:Qwen/Qwen3-Coder-Next", "base_model:quantized:Qwen/Qwen3-Coder-Next", "license:apache-2.0", "8-bit", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_next (Qwen3NextForCausalLM)", "total_params_b": 79.674391296, "active_params_b": 3.56376448, "layers": 48, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "MLX 8-bit affine", "weight_bytes_per_param": 1.0625163716343329, "resident_weight_gb": 84.655345152, "active_weight_gb": 3.78780416, "kv_alloc_gb_per_1k_tokens": 0.024576, "kv_read_gb_per_1k_tokens": 0.024576, "kv_alloc_fixed_gb": 0.077856768, "kv_read_fixed_gb": 0.077856768, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, pinned served MLX config, official BF16 base config comparison, model card, safetensors index, direct shard header range reads, and the Transformers qwen3_next implementation. Bound weights use the exact 84.655345152 GB resident MLX package and 3.787804160 GB single-session ordinary text-decode traffic; the audited profile supersedes the generated ideal 8-bit estimate, fixes shared expert count, and adds the fixed DeltaNet runtime state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-Next-MLX-8bit" }, { "label": "Pinned Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3-Coder-Next-MLX-8bit" }, { "label": "Pinned model config", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-Next-MLX-8bit/raw/f911d6c6cb4cf62fd3c5fd4909feb046872e0221/config.json" }, { "label": "Pinned BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3-Coder-Next/raw/a7fbcb5c0e12d62a448eaa0e260346bf5dcc0feb/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-Next-MLX-8bit/raw/f911d6c6cb4cf62fd3c5fd4909feb046872e0221/model.safetensors.index.json" }, { "label": "Transformers Qwen3-Next implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/qwen3_next/modeling_qwen3_next.py" } ] }, { "id": "meta-llama/Llama-3.1-405B-Instruct", "name": "Llama-3.1-405B-Instruct BF16", "short_name": "Llama 3.1 405B", "license": "llama3.1", "hf_pipeline_tag": "text-generation", "hf_downloads": 210625, "tags": [ "transformers", "safetensors", "llama", "text-generation", "facebook", "meta", "pytorch", "llama-3", "conversational", "en", "de", "fr", "it", "pt", "hi", "es", "th", "arxiv:2204.05149", "base_model:meta-llama/Llama-3.1-405B", "base_model:finetune:meta-llama/Llama-3.1-405B", "license:llama3.1", "eval-results", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Unknown gated model", "detail": "gated config unavailable", "total_params_b": 405.8533888, "active_params_b": 405.8533888, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 811.7067776, "active_weight_gb": 811.7067776, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Fail-closed on 2026-07-06 from current Hugging Face API metadata and access checks. The API records 405.853388800B BF16 safetensors parameters at commit be673f326cab4cd22ccfef76109faf68e41aa5f1, or 811.706777600 GB if charged by dtype. Raw config, model.safetensors.index.json, a safetensors shard HEAD request, and hf download checks require manual approval in this environment. KV geometry, max context, tied embeddings, and swept decode traffic are intentionally unknown until direct gated evidence is available." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/meta-llama/Llama-3.1-405B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.1-405B-Instruct" }, { "label": "Base model card", "url": "https://huggingface.co/meta-llama/Llama-3.1-405B" }, { "label": "Gated config access check", "url": "https://huggingface.co/meta-llama/Llama-3.1-405B-Instruct/raw/be673f326cab4cd22ccfef76109faf68e41aa5f1/config.json" }, { "label": "Gated tensor index access check", "url": "https://huggingface.co/meta-llama/Llama-3.1-405B-Instruct/raw/be673f326cab4cd22ccfef76109faf68e41aa5f1/model.safetensors.index.json" } ] }, { "id": "baidu/ERNIE-4.5-VL-28B-A3B-PT", "name": "ERNIE-4.5-VL-28B-A3B-PT F32/BF16/F16", "short_name": "ERNIE 4.5 VL 28B A3B PT", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 215890, "tags": [ "transformers", "safetensors", "ernie4_5_moe_vl", "image-text-to-text", "ERNIE4.5", "conversational", "custom_code", "en", "zh", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "ernie4_5_moe_vl (Ernie4_5_VLMoeForConditionalGeneration)", "total_params_b": 29.399294336, "active_params_b": 3, "layers": 28, "attention_heads": 20, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": 10, "routed_experts_per_token": 1, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "F32/BF16/F16", "weight_bytes_per_param": 2.0006021107785, "resident_weight_gb": 58.816290304, "active_weight_gb": 6.713158144, "kv_alloc_gb_per_1k_tokens": 0.057344, "kv_read_gb_per_1k_tokens": 0.057344, "notes": "Audited on 2026-07-06 from current Hugging Face API metadata, pinned config/modeling/index files, and direct safetensors header range reads. Resident tensor payloads sum exactly to 58.816290304 GB across F32 0.035403264 GB, BF16 57.520020480 GB, and F16 1.260866560 GB. Ordinary text-decode traffic charges the tied output embedding, fixed text path, shared experts, and six selected text experts for 6.713158144 GB. Vision encoder, resampler, second 64-expert vision pool, and second multimodal gate are resident-only for ordinary text decode." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/baidu/ERNIE-4.5-VL-28B-A3B-PT" }, { "label": "Model config", "url": "https://huggingface.co/baidu/ERNIE-4.5-VL-28B-A3B-PT/raw/e3815e65c607ea211bfe21b46ab0cd264b76731c/config.json" }, { "label": "Custom runtime implementation", "url": "https://huggingface.co/baidu/ERNIE-4.5-VL-28B-A3B-PT/raw/e3815e65c607ea211bfe21b46ab0cd264b76731c/modeling_ernie4_5_vl.py" }, { "label": "Safetensors index", "url": "https://huggingface.co/baidu/ERNIE-4.5-VL-28B-A3B-PT/raw/e3815e65c607ea211bfe21b46ab0cd264b76731c/model.safetensors.index.json" } ] }, { "id": "ykarout/Qwen3.6-35B-A3B-NVFP4", "name": "Qwen3.6-35B-A3B-NVFP4 NVFP4", "short_name": "Qwen3.6 35B A3B NVFP4", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 202525, "tags": [ "transformers", "safetensors", "qwen3_5_moe", "image-text-to-text", "NVFP4", "4-bit", "FP4", "Qwen3.6-35B-A3B-NVFP4", "Qwen3.6-35B-A3B-FP4", "Blackwell", "Qwen3.6", "conversational", "base_model:Qwen/Qwen3.6-35B-A3B", "base_model:quantized:Qwen/Qwen3.6-35B-A3B", "license:apache-2.0", "endpoints_compatible", "modelopt", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_5_moe_text", "total_params_b": 35.9518, "active_params_b": 3, "layers": 40, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "NVFP4", "weight_bytes_per_param": 0.6650395972385249, "resident_weight_gb": 23.909370592, "active_weight_gb": 4.445828096, "kv_alloc_gb_per_1k_tokens": 0.01024, "kv_read_gb_per_1k_tokens": 0.01024, "notes": "Audited on 2026-07-06 from current Hugging Face API metadata, pinned config/hf_quant_config/index files, current base config comparison, and direct safetensors shard-header range reads. Resident tensor payloads sum exactly to 23.909370592 GB. Ordinary text-decode traffic charges lm_head, DeltaNet weights, full-attention weights, shared experts, gates, norms, and eight selected routed experts for 4.445828096 GB. The simple KV coefficient records the FP8 full-attention stream only; the self-contained profile also charges the fixed DeltaNet runtime state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/ykarout/Qwen3.6-35B-A3B-NVFP4" }, { "label": "Model config", "url": "https://huggingface.co/ykarout/Qwen3.6-35B-A3B-NVFP4/raw/5093c2a1d4d263007b563753438caf75dadaec48/config.json" }, { "label": "ModelOpt quantization config", "url": "https://huggingface.co/ykarout/Qwen3.6-35B-A3B-NVFP4/raw/5093c2a1d4d263007b563753438caf75dadaec48/hf_quant_config.json" }, { "label": "Safetensors index", "url": "https://huggingface.co/ykarout/Qwen3.6-35B-A3B-NVFP4/raw/5093c2a1d4d263007b563753438caf75dadaec48/model.safetensors.index.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B" } ] }, { "id": "google/medgemma-4b-it", "name": "medgemma-4b-it BF16", "short_name": "medgemma 4b it", "license": "other", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 291353, "tags": [ "transformers", "safetensors", "gemma3", "image-text-to-text", "medical", "radiology", "clinical-reasoning", "dermatology", "pathology", "ophthalmology", "chest-x-ray", "conversational", "arxiv:2303.15343", "arxiv:2507.05201", "arxiv:2405.03162", "arxiv:2106.14463", "arxiv:2412.03555", "arxiv:2501.19393", "arxiv:2009.13081", "arxiv:2102.09542", "arxiv:2411.15640", "arxiv:2404.05590", "arxiv:2501.18362", "base_model:google/medgemma-4b-pt", "base_model:finetune:google/medgemma-4b-pt", "license:other", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense multimodal transformer", "detail": "Gemma3ForConditionalGeneration (gated raw config)", "total_params_b": 4.300079472, "active_params_b": 4.300079472, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 8.600158944, "active_weight_gb": 8.600158944, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manually corrected on 2026-07-06 from current Hugging Face API metadata. The repo is auto-gated; raw config, README, safetensors index, and hf download config.json all returned access denied. Production bounds use the unsupported profile and must not infer KV geometry from this row." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/medgemma-4b-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/medgemma-4b-it" } ] }, { "id": "Qwen/Qwen3Guard-Gen-0.6B", "name": "Qwen3Guard-Gen-0.6B BF16", "short_name": "Qwen3Guard Gen 0.6B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 185541, "tags": [ "transformers", "safetensors", "qwen3", "text-generation", "conversational", "arxiv:2510.14276", "base_model:Qwen/Qwen3-0.6B", "base_model:finetune:Qwen/Qwen3-0.6B", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 0.751632384, "active_params_b": 0.59604992, "layers": 28, "attention_heads": 16, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 1.503264768, "active_weight_gb": 1.19209984, "kv_alloc_gb_per_1k_tokens": 0.114688, "kv_read_gb_per_1k_tokens": 0.114688, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, pinned served config, model card, audited Qwen3 0.6B profile comparison, and a direct safetensors header range read. Resident bytes are exact BF16 tensor payloads. Ordinary text-decode traffic excludes the input embedding lookup and includes the separately stored lm_head.weight output projection." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3Guard-Gen-0.6B" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3Guard-Gen-0.6B/raw/fada3b2f655b89601929198343c94cd2f64d93cc/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3Guard-Gen-0.6B" }, { "label": "Safetensors artifact", "url": "https://huggingface.co/Qwen/Qwen3Guard-Gen-0.6B/resolve/fada3b2f655b89601929198343c94cd2f64d93cc/model.safetensors" } ] }, { "id": "XiaomiMiMo/MiMo-7B-Base", "name": "MiMo-7B-Base BF16", "short_name": "MiMo 7B", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 242871, "tags": [ "transformers", "safetensors", "mimo", "text-generation", "conversational", "custom_code", "arxiv:2505.07608", "license:mit", "eval-results", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "mimo (MiMoForCausalLM)", "total_params_b": 7.833409536, "active_params_b": 7.001337856, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 15.666819072, "active_weight_gb": 14.002675712, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, pinned model card, served config, generation config, custom modeling code, safetensors index, and direct safetensors shard header range reads. Exact tensor payload bytes: 15.666819072 GB resident, 14.002675712 GB ordinary text-decode swept traffic, and 1.664143360 GB resident-only input embedding plus MTP tensors. This catalog row models ordinary target-model decode; MiMo-MTP speculative serving requires a separate adapter." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/XiaomiMiMo/MiMo-7B-Base" }, { "label": "Model config", "url": "https://huggingface.co/XiaomiMiMo/MiMo-7B-Base/raw/c72df4586cb8bdeebd65f36929cd3385a6566fbe/config.json" }, { "label": "Generation config", "url": "https://huggingface.co/XiaomiMiMo/MiMo-7B-Base/raw/c72df4586cb8bdeebd65f36929cd3385a6566fbe/generation_config.json" }, { "label": "Custom modeling code", "url": "https://huggingface.co/XiaomiMiMo/MiMo-7B-Base/raw/c72df4586cb8bdeebd65f36929cd3385a6566fbe/modeling_mimo.py" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/XiaomiMiMo/MiMo-7B-Base" }, { "label": "Safetensors index", "url": "https://huggingface.co/XiaomiMiMo/MiMo-7B-Base/raw/c72df4586cb8bdeebd65f36929cd3385a6566fbe/model.safetensors.index.json" } ] }, { "id": "stepfun-ai/Step-3.5-Flash", "name": "Step-3.5-Flash BF16/F32", "short_name": "Step 3.5 Flash", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 178273, "tags": [ "transformers", "safetensors", "step3p5", "text-generation", "conversational", "custom_code", "arxiv:2602.10604", "arxiv:2601.05593", "arxiv:2507.19427", "license:apache-2.0", "eval-results", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "step3p5 (Step3p5ForCausalLM)", "total_params_b": 199.384301376, "active_params_b": 11.459378048, "layers": 45, "attention_heads": 64, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 288, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "BF16/F32", "weight_bytes_per_param": 2.0000001213335246, "resident_weight_gb": 398.768626944, "active_weight_gb": 22.918756096, "kv_alloc_gb_per_1k_tokens": 0.049152, "kv_read_gb_per_1k_tokens": 0.049152, "kv_alloc_fixed_gb": 0.069206016, "kv_read_fixed_gb": 0.069206016, "expert_param_b": 0.66060288, "shared_expert_param_b": 0, "fixed_param_b": 6.174555008, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, model card, pinned served config, remote modeling code, safetensors index, and direct shard-header range reads. Bound weights use exact stored bytes: 398.768626944 GB resident package, 12.349110016 GB fixed ordinary traffic, 1.321205760 GB per routed expert index across 288 experts, and 5.912258048 GB resident-only input embedding plus MTP tensors. KV traffic uses BF16 full/sliding cache: 12 full-context ordinary layers plus 33 sliding-window ordinary layers capped at 512 tokens." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/stepfun-ai/Step-3.5-Flash" }, { "label": "Model config", "url": "https://huggingface.co/stepfun-ai/Step-3.5-Flash/raw/ab446a3de5e171ea341227e24bb1f090e1b771f7/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/stepfun-ai/Step-3.5-Flash" }, { "label": "Custom modeling code", "url": "https://huggingface.co/stepfun-ai/Step-3.5-Flash/raw/ab446a3de5e171ea341227e24bb1f090e1b771f7/modeling_step3p5.py" }, { "label": "Safetensors index and shard-header audit", "url": "https://huggingface.co/stepfun-ai/Step-3.5-Flash/resolve/ab446a3de5e171ea341227e24bb1f090e1b771f7/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen1.5-MoE-A2.7B", "name": "Qwen1.5-MoE-A2.7B BF16", "short_name": "Qwen1.5 MoE A2.7B", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 216320, "tags": [ "transformers", "safetensors", "qwen2_moe", "text-generation", "pretrained", "moe", "conversational", "en", "license:other", "endpoints_compatible", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen2_moe (Qwen2MoeForCausalLM)", "total_params_b": 14.315784192, "active_params_b": 2.7, "layers": 24, "attention_heads": 16, "kv_heads": 16, "head_dim": 128, "max_context_tokens": 8192, "routed_experts": 60, "routed_experts_per_token": 4, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 28.631568384, "active_weight_gb": 4.756017152, "kv_alloc_gb_per_1k_tokens": 0.196608, "kv_read_gb_per_1k_tokens": 0.196608, "expert_param_b": 0.207618048, "shared_expert_param_b": 0.830521344, "fixed_param_b": 1.547536384, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, model card, pinned served config, pinned Transformers implementation, safetensors index, and direct shard-header range reads. Catalog parameters keep the model-card 2.7B active value, while bound weights use exact stored bytes: 28.631568384 GB resident, 3.095072768 GB fixed ordinary traffic including the always-on shared expert path, 0.415236096 GB per routed expert index across 60 experts, and 0.622329856 GB resident-only input embeddings." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen1.5-MoE-A2.7B" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen1.5-MoE-A2.7B/raw/1a758c50ecb6350748b9ce0a99d2352fd9fc11c9/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen1.5-MoE-A2.7B" }, { "label": "Transformers Qwen2-MoE implementation", "url": "https://raw.githubusercontent.com/huggingface/transformers/0b05eeced8e8e4a5b2682000a8c1b151e76a9181/src/transformers/models/qwen2_moe/modeling_qwen2_moe.py" }, { "label": "Safetensors index and shard-header audit", "url": "https://huggingface.co/Qwen/Qwen1.5-MoE-A2.7B/resolve/1a758c50ecb6350748b9ce0a99d2352fd9fc11c9/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen3-4B-Thinking-2507-FP8", "name": "Qwen3-4B-Thinking-2507-FP8 FP8", "short_name": "Qwen3 4B Thinking 2507 FP8", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 207284, "tags": [ "transformers", "safetensors", "qwen3", "text-generation", "conversational", "arxiv:2505.09388", "base_model:Qwen/Qwen3-4B-Thinking-2507", "base_model:quantized:Qwen/Qwen3-4B-Thinking-2507", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "fp8", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 4.411646016, "active_params_b": 4.022689856, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "FP8", "weight_bytes_per_param": 1, "resident_weight_gb": 5.189976192, "active_weight_gb": 4.412063872, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, model card, served FP8 config, audited BF16 base profile comparison, and direct safetensors header range read. Bound weights separate the 5.189976192 GB resident FP8/F16 package from the 4.412063872 GB ordinary text-decode swept subset; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-4B-Thinking-2507-FP8" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-4B-Thinking-2507-FP8/raw/953532f942706930ec4bb870569932ef63038fdf/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-4B-Thinking-2507" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-4B-Thinking-2507-FP8" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/Qwen/Qwen3-4B-Thinking-2507-FP8/resolve/953532f942706930ec4bb870569932ef63038fdf/model.safetensors" } ] }, { "id": "Qwen/Qwen3-30B-A3B-GPTQ-Int4", "name": "Qwen3-30B-A3B-GPTQ-Int4 4-bit", "short_name": "Qwen3 30B A3B GPTQ Int4", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 136678, "tags": [ "transformers", "safetensors", "qwen3_moe", "text-generation", "conversational", "arxiv:2309.00071", "arxiv:2505.09388", "base_model:Qwen/Qwen3-30B-A3B", "base_model:quantized:Qwen/Qwen3-30B-A3B", "license:apache-2.0", "endpoints_compatible", "4-bit", "gptq", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_moe (Qwen3MoeForCausalLM)", "total_params_b": 30.532122624, "active_params_b": 3.3, "layers": 48, "attention_heads": 32, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "GPTQ Int4 + F16 side tensors", "weight_bytes_per_param": 0.5543072321705084, "resident_weight_gb": 16.924176384, "active_weight_gb": 2.069393408, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "expert_param_b": 0.21396757811652858, "fixed_param_b": 2.021556160492793, "shared_expert_param_b": 0, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, model card, pinned served GPTQ config, quantize_config.json, BF16 base config comparison, and direct safetensors header range reads. Bound weights separate the 16.924176384 GB resident GPTQ/F16 package from 2.069393408 GB ordinary single-session text-decode swept traffic: 1.120563200 GB fixed traffic plus 8 routed expert groups at 0.118603776 GB each. The audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-GPTQ-Int4" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-GPTQ-Int4/raw/9b534e4318b7ebc3c961a839f13eb18b1833f441/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-30B-A3B-GPTQ-Int4" }, { "label": "Quantize config", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-GPTQ-Int4/raw/9b534e4318b7ebc3c961a839f13eb18b1833f441/quantize_config.json" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/Qwen/Qwen3-30B-A3B-GPTQ-Int4/resolve/9b534e4318b7ebc3c961a839f13eb18b1833f441/model.safetensors" } ] }, { "id": "microsoft/Phi-mini-MoE-instruct", "name": "Phi-mini-MoE-instruct BF16", "short_name": "Phi mini MoE", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 196717, "tags": [ "transformers", "safetensors", "phimoe", "text-generation", "conversational", "custom_code", "en", "arxiv:2506.18349", "arxiv:2404.14219", "arxiv:2409.12136", "license:mit", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "phimoe (PhiMoEForCausalLM)", "total_params_b": 7.647632704, "active_params_b": 2.4, "layers": 32, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 4096, "routed_experts": 16, "routed_experts_per_token": 2, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 15.295265408, "active_weight_gb": 4.46295104, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "expert_param_b": 0.37748736, "fixed_param_b": 1.4765008, "shared_expert_param_b": 0, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, model card, pinned config, pinned SlimMoE runtime code, safetensors index, and direct safetensors header range reads. Bound weights separate the 15.295265408 GB resident BF16 package from 4.462951040 GB ordinary single-session text-decode swept traffic: 2.953001600 GB fixed traffic plus 2 routed expert groups at 0.754974720 GB each. The authoritative profile caps KV traffic at the 2047-token sliding window; these legacy per-1k KV coefficients remain uncapped fallback display fields." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/microsoft/Phi-mini-MoE-instruct" }, { "label": "Model config", "url": "https://huggingface.co/microsoft/Phi-mini-MoE-instruct/raw/f620b32c0d3e8f7e76f57ccdaa88e0df8bc8bfcd/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/microsoft/Phi-mini-MoE-instruct" }, { "label": "Custom SlimMoE runtime code", "url": "https://huggingface.co/microsoft/Phi-mini-MoE-instruct/raw/f620b32c0d3e8f7e76f57ccdaa88e0df8bc8bfcd/modeling_slimmoe.py" }, { "label": "Safetensors index", "url": "https://huggingface.co/microsoft/Phi-mini-MoE-instruct/resolve/f620b32c0d3e8f7e76f57ccdaa88e0df8bc8bfcd/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen2.5-0.5B-Instruct-GGUF", "name": "Qwen2.5-0.5B-Instruct-GGUF GGUF", "short_name": "Qwen2.5 0.5B Instruct", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 148822, "tags": [ "gguf", "chat", "text-generation", "en", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-0.5B-Instruct", "base_model:quantized:Qwen/Qwen2.5-0.5B-Instruct", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 GGUF selected FP16 artifact", "total_params_b": 0.630167424, "active_params_b": 0.494032768, "layers": 24, "attention_heads": 14, "kv_heads": 2, "head_dim": 64, "max_context_tokens": 8192, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF FP16/F32 selected file", "weight_bytes_per_param": 2, "resident_weight_gb": 1.266425696, "active_weight_gb": 0.98820864, "kv_alloc_gb_per_1k_tokens": 0.012288, "kv_read_gb_per_1k_tokens": 0.012288, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, model card, audited BF16 base profile comparison, linked GGUF file sizes, and direct selected FP16 GGUF tensor-index range read. The authoritative bounds profile uses qwen2.5-0.5b-instruct-fp16.gguf because it matches API gguf.totalFileSize: 1.266425696 GB resident and 0.988208640 GB ordinary text-decode swept traffic after excluding token_embd.weight. The selected GGUF header records 8192 context, while the card and BF16 base record 32768 full context with 8192 generation." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-0.5B-Instruct-GGUF" }, { "label": "Selected FP16 GGUF artifact", "url": "https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct-GGUF/resolve/9217f5db79a29953eb74d5343926648285ec7e67/qwen2.5-0.5b-instruct-fp16.gguf" } ] }, { "id": "TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF", "name": "TinyLlama-1.1B-Chat-v1.0-GGUF GGUF", "short_name": "TinyLlama 1.1B Chat v1.0", "license": "apache-2.0", "hf_pipeline_tag": "unknown", "hf_downloads": 194341, "tags": [ "transformers", "gguf", "tinyllama", "en", "dataset:cerebras/SlimPajama-627B", "dataset:bigcode/starcoderdata", "dataset:OpenAssistant/oasst_top1_2023-08-25", "base_model:TinyLlama/TinyLlama-1.1B-Chat-v1.0", "base_model:quantized:TinyLlama/TinyLlama-1.1B-Chat-v1.0", "license:apache-2.0", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "llama GGUF selected Q4_K_M artifact", "total_params_b": 1.100048384, "active_params_b": 1.034512384, "layers": 22, "attention_heads": 32, "kv_heads": 4, "head_dim": 64, "max_context_tokens": 2048, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q4_K_M", "weight_bytes_per_param": 0.6079624366776943, "resident_weight_gb": 0.668788096, "active_weight_gb": 0.630214656, "kv_alloc_gb_per_1k_tokens": 0.022528, "kv_read_gb_per_1k_tokens": 0.022528, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, pinned model card, pinned TinyLlama base config, linked GGUF file sizes, and direct Q4_K_M GGUF header range read. The authoritative profile uses the card-selected tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf artifact: 0.668788096 GB resident, 0.630214656 GB swept ordinary text traffic, 0.038573440 GB resident-only token embedding plus GGUF metadata/header/file overhead, and full-context FP16 KV. The HF API gguf.totalFileSize points to Q2_K, so the selected artifact is intentionally pinned by the model card examples." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF/raw/52e7645ba7c309695bec7ac98f4f005b139cf465/README.md" }, { "label": "Base model config", "url": "https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v1.0/raw/fe8a4ea1ffedaf415f4da2f062534de366a451e6/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v1.0" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF" }, { "label": "Selected Q4_K_M GGUF header audit", "url": "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF/resolve/52e7645ba7c309695bec7ac98f4f005b139cf465/tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf" } ] }, { "id": "Qwen/Qwen2.5-Coder-32B-Instruct-GPTQ-Int4", "name": "Qwen2.5-Coder-32B-Instruct-GPTQ-Int4 4-bit", "short_name": "Qwen2.5 Coder 32B Instruct GPTQ Int4", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 148914, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "code", "codeqwen", "chat", "qwen", "qwen-coder", "conversational", "en", "arxiv:2409.12186", "arxiv:2309.00071", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-Coder-32B-Instruct", "base_model:quantized:Qwen/Qwen2.5-Coder-32B-Instruct", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "4-bit", "gptq", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM, GPTQ Int4)", "total_params_b": 32.763876352, "active_params_b": 31.985308672, "layers": 64, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GPTQ Int4 + F16 side tensors", "weight_bytes_per_param": 0.590398610475137, "resident_weight_gb": 19.343747072, "active_weight_gb": 17.786611712, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, pinned served GPTQ config, model card, audited BF16 Coder base profile comparison, safetensors index, and direct shard-header range reads. Bound weights separate the exact 19.343747072 GB resident GPTQ/F16 package from 17.786611712 GB ordinary text-decode swept traffic; model.embed_tokens.weight remains resident-only. The audited profile supersedes the generated ideal 0.5-byte estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct-GPTQ-Int4/blob/5fece1e232475f4da75cf0e4283606a3186fee8a/README.md" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct-GPTQ-Int4/raw/5fece1e232475f4da75cf0e4283606a3186fee8a/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Coder-32B-Instruct-GPTQ-Int4" }, { "label": "Safetensors index", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct-GPTQ-Int4/resolve/5fece1e232475f4da75cf0e4283606a3186fee8a/model.safetensors.index.json" } ] }, { "id": "cyankiwi/Qwen3-Next-80B-A3B-Instruct-AWQ-4bit", "name": "Qwen3-Next-80B-A3B-Instruct-AWQ-4bit 4-bit", "short_name": "Qwen3 Next 80B A3B Instruct AWQ 4bit", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 195424, "tags": [ "transformers", "safetensors", "qwen3_next", "text-generation", "conversational", "arxiv:2309.00071", "arxiv:2404.06654", "arxiv:2505.09388", "arxiv:2501.15383", "base_model:Qwen/Qwen3-Next-80B-A3B-Instruct", "base_model:quantized:Qwen/Qwen3-Next-80B-A3B-Instruct", "license:apache-2.0", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_next (Qwen3NextForCausalLM, compressed-tensors AWQ)", "total_params_b": 83.791264, "active_params_b": 3, "layers": 48, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "compressed-tensors AWQ Int4 packed I32 + BF16/I64 side tensors", "weight_bytes_per_param": 0.5872139018215551, "resident_weight_gb": 49.203395072, "active_weight_gb": 4.95699968, "kv_alloc_gb_per_1k_tokens": 0.024576, "kv_read_gb_per_1k_tokens": 0.024576, "kv_alloc_fixed_gb": 0.077856768, "kv_read_fixed_gb": 0.077856768, "expert_param_b": 0.14464398703185163, "fixed_param_b": 6.995117226036387, "shared_expert_param_b": 0, "notes": "Manually corrected on 2026-07-06 from the scraped catalog row, current Hugging Face API metadata, pinned compressed-tensors config, BF16 base profile/config comparison, recipe.yaml, safetensors index, and direct range-read safetensors shard headers. Legacy fallback values mirror the audited profile: exact stored resident bytes, fixed always-on hybrid/shared-expert traffic, expected distinct routed experts, BF16 attention KV, and fixed BF16-convolution/F32-SSM DeltaNet recurrent state. The authoritative profile supersedes the generated 41.8957 GB resident estimate and the missing shared-expert count." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/cyankiwi/Qwen3-Next-80B-A3B-Instruct-AWQ-4bit" }, { "label": "Model config", "url": "https://huggingface.co/cyankiwi/Qwen3-Next-80B-A3B-Instruct-AWQ-4bit/raw/b213ae173e2ec9eebe9191b0ee2686c3173d520c/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3-Next-80B-A3B-Instruct-AWQ-4bit" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/cyankiwi/Qwen3-Next-80B-A3B-Instruct-AWQ-4bit/resolve/b213ae173e2ec9eebe9191b0ee2686c3173d520c/model.safetensors.index.json" }, { "label": "Transformers Qwen3-Next implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/qwen3_next/modeling_qwen3_next.py" } ] }, { "id": "abhishekchohan/gemma-3-12b-it-quantized-W4A16", "name": "gemma-3-12b-it-quantized-W4A16 I64/I32/BF16", "short_name": "gemma 3 12b it quantized W4A16", "license": "gemma", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 195261, "tags": [ "transformers", "safetensors", "gemma3", "image-text-to-text", "conversational", "base_model:google/gemma-3-12b-it", "base_model:quantized:google/gemma-3-12b-it", "license:gemma", "text-generation-inference", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense multimodal transformer", "detail": "gemma3_text (Gemma3ForConditionalGeneration, compressed-tensors W4A16)", "total_params_b": 12.271375632, "active_params_b": 11.850084768, "layers": 48, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "compressed-tensors W4A16", "weight_bytes_per_param": 0.6849411922557306, "resident_weight_gb": 8.405170656, "active_weight_gb": 7.562588928, "kv_alloc_gb_per_1k_tokens": 0.065536, "kv_read_gb_per_1k_tokens": 0.065536, "kv_alloc_fixed_gb": 0.33554432, "kv_read_fixed_gb": 0.33554432, "notes": "Manually audited on 2026-07-06 from current HF API metadata, served compressed-tensors config, model card metadata, quantization recipe, safetensors index, direct safetensors header range reads, and Gemma 3 local/global attention documentation. downloads=195261. Exact tensor payload bytes: 8.405170656 GB resident, 7.562588928 GB swept ordinary text traffic including the tied input embedding/output projection, and 0.842581728 GB resident-only vision/projector bytes. KV fields split 8 full-context layers per 1K tokens plus 40 sliding-window layers capped at 1024 tokens." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/abhishekchohan/gemma-3-12b-it-quantized-W4A16" }, { "label": "Model config", "url": "https://huggingface.co/abhishekchohan/gemma-3-12b-it-quantized-W4A16/raw/d9833e9131576955f818012055a0bc9b74b46e64/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-3-12b-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/abhishekchohan/gemma-3-12b-it-quantized-W4A16" }, { "label": "Quantization recipe", "url": "https://huggingface.co/abhishekchohan/gemma-3-12b-it-quantized-W4A16/raw/d9833e9131576955f818012055a0bc9b74b46e64/recipe.yaml" }, { "label": "Safetensors index", "url": "https://huggingface.co/abhishekchohan/gemma-3-12b-it-quantized-W4A16/resolve/d9833e9131576955f818012055a0bc9b74b46e64/model.safetensors.index.json" } ] }, { "id": "decart-ai/Kimi-K2.7-Code-NVFP4", "name": "Kimi-K2.7-Code-NVFP4 NVFP4", "short_name": "Kimi K2.7 Code NVFP4", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 352463, "tags": [ "Model Optimizer", "safetensors", "kimi_k25", "ModelOpt", "Kimi-K2.7-Code", "quantized", "FP4", "fp4", "NVFP4", "text-generation", "conversational", "custom_code", "base_model:moonshotai/Kimi-K2.7-Code", "base_model:quantized:moonshotai/Kimi-K2.7-Code", "license:other", "modelopt", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "kimi_k25 (KimiK25ForConditionalGeneration / DeepseekV3ForCausalLM, ModelOpt NVFP4)", "total_params_b": 1058.589420528, "active_params_b": 32, "layers": 61, "attention_heads": 64, "kv_heads": 64, "head_dim": 112, "max_context_tokens": 262144, "routed_experts": 384, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "NVFP4/BF16/FP8 ModelOpt", "weight_bytes_per_param": 0.5, "resident_weight_gb": 595.148146656, "active_weight_gb": 32.986470656, "kv_alloc_gb_per_1k_tokens": 1.24928, "kv_read_gb_per_1k_tokens": 1.24928, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, Decart model card, served config, hf_quant_config, base-config comparison, custom-code hash comparison, safetensors index, and direct shard header range reads. Bound weights separate the exact 595.148146656 GB resident ModelOpt package from 32.986470656 GB single-session ordinary text-decode MoE traffic and charge expanded FP8 K/V cache streams for ordinary text decode." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/decart-ai/Kimi-K2.7-Code-NVFP4" }, { "label": "Model config", "url": "https://huggingface.co/decart-ai/Kimi-K2.7-Code-NVFP4/raw/b16b4af9d9255c114bf3d5f124dc1ba12cbf812c/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/moonshotai/Kimi-K2.7-Code" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/decart-ai/Kimi-K2.7-Code-NVFP4" }, { "label": "ModelOpt quantization config", "url": "https://huggingface.co/decart-ai/Kimi-K2.7-Code-NVFP4/raw/b16b4af9d9255c114bf3d5f124dc1ba12cbf812c/hf_quant_config.json" }, { "label": "Safetensors index", "url": "https://huggingface.co/decart-ai/Kimi-K2.7-Code-NVFP4/resolve/b16b4af9d9255c114bf3d5f124dc1ba12cbf812c/model.safetensors.index.json" } ] }, { "id": "nvidia/MiniMax-M2.7-NVFP4", "name": "MiniMax-M2.7-NVFP4 NVFP4", "short_name": "MiniMax M2.7 NVFP4", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 100248, "tags": [ "Model Optimizer", "safetensors", "minimax_m2", "nvidia", "ModelOpt", "MiniMax", "quantized", "NVFP4", "nvfp4", "text-generation", "conversational", "custom_code", "base_model:MiniMaxAI/MiniMax-M2.7", "base_model:quantized:MiniMaxAI/MiniMax-M2.7", "license:other", "8-bit", "modelopt", "deploy:azure", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "minimax_m2 (MiniMaxM2ForCausalLM, ModelOpt NVFP4)", "total_params_b": 228.703644928, "active_params_b": 10.41655168, "layers": 62, "attention_heads": 48, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 196608, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "NVFP4/BF16/FP8 ModelOpt", "weight_bytes_per_param": 0.6115456065076326, "resident_weight_gb": 139.862709248, "active_weight_gb": 16.199840384, "kv_alloc_gb_per_1k_tokens": 0.126976, "kv_read_gb_per_1k_tokens": 0.126976, "notes": "Manually audited on 2026-07-06 from HF API commit e79701cb1f9dce8fe5395b9ed2b20170beebecde, model card, served ModelOpt config, hf_quant_config, audited FP8 base profile comparison, custom code hash comparison, safetensors index, and direct safetensors header range reads. Active traffic is fixed decode traffic plus 8 routed expert groups; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/MiniMax-M2.7-NVFP4" }, { "label": "Model config", "url": "https://huggingface.co/nvidia/MiniMax-M2.7-NVFP4/raw/e79701cb1f9dce8fe5395b9ed2b20170beebecde/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M2.7" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/MiniMax-M2.7-NVFP4" }, { "label": "ModelOpt quantization config", "url": "https://huggingface.co/nvidia/MiniMax-M2.7-NVFP4/raw/e79701cb1f9dce8fe5395b9ed2b20170beebecde/hf_quant_config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/nvidia/MiniMax-M2.7-NVFP4/resolve/e79701cb1f9dce8fe5395b9ed2b20170beebecde/model.safetensors.index.json" } ] }, { "id": "OpenGVLab/InternVL2_5-8B-AWQ", "name": "InternVL2_5-8B-AWQ 4-bit", "short_name": "InternVL2 5 8B", "license": "mit", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 304286, "tags": [ "transformers", "pytorch", "internvl_chat", "feature-extraction", "internvl", "custom_code", "image-text-to-text", "conversational", "multilingual", "arxiv:2312.14238", "arxiv:2404.16821", "arxiv:2410.16261", "arxiv:2412.05271", "base_model:OpenGVLab/InternVL2_5-8B", "base_model:quantized:OpenGVLab/InternVL2_5-8B", "license:mit", "region:us" ], "architecture": { "type": "Dense multimodal transformer", "detail": "internvl_chat (InternVLChatModel, InternLM2ForCausalLM AWQ)", "total_params_b": 8.075365376, "active_params_b": 7.358685184, "layers": 32, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "AWQ INT4 + FP16 side tensors", "weight_bytes_per_param": 0.7204705393629981, "resident_weight_gb": 5.818062848, "active_weight_gb": 4.384702464, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "notes": "Manually audited on 2026-07-06 from current Hugging Face 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. Bound weights separate the exact 5.818062848 GB resident AWQ/FP16 multimodal package from the 4.384702464 GB ordinary text-decode swept subset; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/OpenGVLab/InternVL2_5-8B-AWQ" }, { "label": "Model config", "url": "https://huggingface.co/OpenGVLab/InternVL2_5-8B-AWQ/raw/14e0a7932fa4d629e19209f7dd1d4003690e2b64/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/OpenGVLab/InternVL2_5-8B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/OpenGVLab/InternVL2_5-8B-AWQ" }, { "label": "Pinned BF16 base config", "url": "https://huggingface.co/OpenGVLab/InternVL2_5-8B/raw/e9e4c0dc1db56bfab10458671519b7fa3dd29463/config.json" }, { "label": "PyTorch checkpoint index and metadata audit", "url": "https://huggingface.co/OpenGVLab/InternVL2_5-8B-AWQ/raw/14e0a7932fa4d629e19209f7dd1d4003690e2b64/pytorch_model.bin.index.json" } ] }, { "id": "Qwen/Qwen2-1.5B", "name": "Qwen2-1.5B BF16", "short_name": "Qwen2 1.5B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 197198, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "pretrained", "conversational", "en", "license:apache-2.0", "eval-results", "text-generation-inference", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 1.543714304, "active_params_b": 1.543714304, "layers": 28, "attention_heads": 12, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 3.087428608, "active_weight_gb": 3.087428608, "kv_alloc_gb_per_1k_tokens": 0.028672, "kv_read_gb_per_1k_tokens": 0.028672, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, model card, served config, direct safetensors header range read, and comparison with the audited Qwen2 1.5B Instruct sibling. The checkpoint has tied embeddings and no separate lm_head.weight, so the embedding/output table remains in ordinary text-decode swept traffic." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2-1.5B" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2-1.5B/raw/8a16abf2848eda07cc5253dec660bf1ce007ad7a/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2-1.5B" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/Qwen/Qwen2-1.5B/resolve/8a16abf2848eda07cc5253dec660bf1ce007ad7a/model.safetensors" } ] }, { "id": "MiniMaxAI/MiniMax-M3", "name": "MiniMax-M3 BF16/F32", "short_name": "MiniMax M3", "license": "other", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 226766, "tags": [ "transformers", "safetensors", "minimax_m3_vl", "image-text-to-text", "multimodal", "moe", "agent", "coding", "video", "conversational", "custom_code", "arxiv:2606.13392", "license:other", "eval-results", "endpoints_compatible", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "minimax_m3_vl (MiniMaxM3SparseForCausalLM)", "total_params_b": 427.04014016, "active_params_b": 23, "layers": 60, "attention_heads": 64, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 1048576, "routed_experts": 128, "routed_experts_per_token": 4, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "BF16/F32", "weight_bytes_per_param": 2.0002170250318043, "resident_weight_gb": 854.17295872, "active_weight_gb": 49.55622912, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manual correction on 2026-07-06 from live Hugging Face API metadata, served BF16 config, model card MSA notes, safetensors index, and direct shard-header range reads. The model card says about 23B activated parameters; active_weight_gb records exact single-token ordinary text weight traffic from headers: 23.736093696 GB fixed traffic plus 4 routed expert groups at 6.455033856 GB each. KV fields are zeroed in this legacy row because MiniMax Sparse Attention allocation/read traffic is not audited; the authoritative profile is fail-closed and returns unsupported_profile for production bounds." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MiniMaxAI/MiniMax-M3" }, { "label": "Model card", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M3" }, { "label": "Pinned served config", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M3/raw/fa6849585c885749b76c0b1f299841b231e4cbfa/config.json" }, { "label": "Safetensors index", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M3/resolve/fa6849585c885749b76c0b1f299841b231e4cbfa/model.safetensors.index.json" } ] }, { "id": "Chunity/gemma-4-E4B-it-AWQ-4bit", "name": "gemma-4-E4B-it-AWQ-4bit 4-bit", "short_name": "gemma 4 E4B it AWQ 4bit", "license": "unknown", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 223205, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "awq", "autoround", "4-bit", "multimodal", "conversational", "base_model:google/gemma-4-E4B-it", "base_model:quantized:google/gemma-4-E4B-it", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 4.805237482, "active_params_b": 1.50857681, "layers": 42, "attention_heads": 8, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "AutoRound AWQ INT4 + BF16/F16 side tensors", "weight_bytes_per_param": 2.189818481067113, "resident_weight_gb": 10.522597844, "active_weight_gb": 3.9292765, "kv_alloc_gb_per_1k_tokens": 0.028672, "kv_read_gb_per_1k_tokens": 0.028672, "kv_alloc_fixed_gb": 0.03670016, "kv_read_fixed_gb": 0.03670016, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, model card, served AutoRound AWQ config, standalone quantization_config, base Google config comparison, linked-object metadata, and direct seven-shard safetensors header range reads. Bound weights separate the exact 10.522597844 GB resident AWQ/BF16/F16 package from the 3.929276500 GB ordinary text-decode swept subset; no lm_head tensor is stored, so the tied embedding/output projection is charged once. Hybrid Gemma 4 E4B KV uses full-context plus fixed sliding-window cache terms. The audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Chunity/gemma-4-E4B-it-AWQ-4bit" }, { "label": "Model config", "url": "https://huggingface.co/Chunity/gemma-4-E4B-it-AWQ-4bit/raw/cae66dfb9e0f442e5ac9d0eab2d9f2ab27024700/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-E4B-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Chunity/gemma-4-E4B-it-AWQ-4bit" }, { "label": "Pinned Google Gemma 4 E4B IT base config", "url": "https://huggingface.co/google/gemma-4-E4B-it/raw/fee6332c1abaafb77f6f9624236c63aa2f1d0187/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Chunity/gemma-4-E4B-it-AWQ-4bit/resolve/cae66dfb9e0f442e5ac9d0eab2d9f2ab27024700/model.safetensors.index.json" } ] }, { "id": "nvidia/music-flamingo-2601-hf", "name": "music-flamingo-2601-hf BF16", "short_name": "music flamingo 2601", "license": "other", "hf_pipeline_tag": "audio-text-to-text", "hf_downloads": 188844, "tags": [ "transformers", "safetensors", "musicflamingo", "text2text-generation", "music/songs", "music understanding", "music reasoning", "audio-text-to-text", "en", "dataset:nvidia/MF-Skills", "arxiv:2511.10289", "arxiv:2505.13032", "license:other", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "musicflamingo / qwen2 text decoder", "total_params_b": 8.26721536, "active_params_b": 7.069214208, "layers": 28, "attention_heads": 28, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 16.53443072, "active_weight_gb": 14.138428416, "kv_alloc_gb_per_1k_tokens": 0.057344, "kv_read_gb_per_1k_tokens": 0.057344, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, model card, served MusicFlamingo config, processor/generation configs, linked-object metadata, and direct safetensors header range read. Bound weights separate the exact 16.534430720 GB resident BF16 package from the 14.138428416 GB ordinary text-decode swept subset; audio_tower, multimodal projector, and input embeddings are resident-only for ordinary text decode. Audio encoder throughput and audio prompt prefill are outside this legacy row; the audited profile supersedes these fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/music-flamingo-2601-hf" }, { "label": "Model config", "url": "https://huggingface.co/nvidia/music-flamingo-2601-hf/raw/6b5be086d52f65a1e204cb0faf70bf54e2741ecd/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/music-flamingo-2601-hf" }, { "label": "Processor config", "url": "https://huggingface.co/nvidia/music-flamingo-2601-hf/raw/6b5be086d52f65a1e204cb0faf70bf54e2741ecd/processor_config.json" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/nvidia/music-flamingo-2601-hf/resolve/6b5be086d52f65a1e204cb0faf70bf54e2741ecd/model.safetensors" } ] }, { "id": "lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-4bit", "name": "Qwen3-Coder-30B-A3B-Instruct-MLX-4bit 4-bit", "short_name": "Qwen3 Coder 30B A3B Instruct MLX 4bit", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 183045, "tags": [ "transformers", "safetensors", "qwen3_moe", "text-generation", "mlx", "conversational", "base_model:Qwen/Qwen3-Coder-30B-A3B-Instruct", "base_model:quantized:Qwen/Qwen3-Coder-30B-A3B-Instruct", "license:apache-2.0", "endpoints_compatible", "4-bit", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_moe (Qwen3MoeForCausalLM)", "total_params_b": 30.532122624, "active_params_b": 3.042353139, "layers": 48, "attention_heads": 32, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "MLX 4-bit affine", "weight_bytes_per_param": 0.5625099315728466, "resident_weight_gb": 17.174622208, "active_weight_gb": 1.711353856, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "notes": "Audited on 2026-07-06 from current Hugging Face API metadata, model card, pinned served MLX config, official BF16 base config comparison, safetensors index, and direct safetensors shard header range reads. Bound weights use exact stored MLX tensor payload bytes: 17.174622208 GB resident, 0.175030272 GB resident-only input embedding, 0.692137984 GB fixed decode traffic, and 0.127401984 GB per routed expert group." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-4bit" }, { "label": "Model config", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-4bit/raw/bbecedefd826819d2c6b6465e88ca7b9b8ad3407/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-4bit" }, { "label": "Safetensors index", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-4bit/raw/bbecedefd826819d2c6b6465e88ca7b9b8ad3407/model.safetensors.index.json" } ] }, { "id": "lmstudio-community/Qwen3-Coder-Next-MLX-4bit", "name": "Qwen3-Coder-Next-MLX-4bit 4-bit", "short_name": "Qwen3 Coder Next MLX 4bit", "license": "apache-2.0", "hf_pipeline_tag": "unknown", "hf_downloads": 181136, "tags": [ "mlx", "safetensors", "qwen3_next", "base_model:Qwen/Qwen3-Coder-Next", "base_model:quantized:Qwen/Qwen3-Coder-Next", "license:apache-2.0", "4-bit", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_next (Qwen3NextForCausalLM)", "total_params_b": 79.674391296, "active_params_b": 3.609954655, "layers": 48, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "MLX 4-bit affine", "weight_bytes_per_param": 0.5628415784615021, "resident_weight_gb": 44.84406016, "active_weight_gb": 2.031832576, "kv_alloc_gb_per_1k_tokens": 0.024576, "kv_read_gb_per_1k_tokens": 0.024576, "kv_alloc_fixed_gb": 0.077856768, "kv_read_fixed_gb": 0.077856768, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, pinned served MLX config, official BF16 base config comparison, model card, safetensors index, direct shard header range reads, and the Transformers qwen3_next implementation. Bound weights use the exact 44.844060160 GB resident MLX package and 2.031832576 GB single-session ordinary text-decode traffic; the audited profile supersedes the generated ideal 4-bit estimate, fixes shared expert count, and adds the fixed DeltaNet runtime state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-Next-MLX-4bit" }, { "label": "Pinned Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3-Coder-Next-MLX-4bit" }, { "label": "Pinned model config", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-Next-MLX-4bit/raw/03cba26036330b7553d252c1f3fb899f16cc5ea5/config.json" }, { "label": "Pinned BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3-Coder-Next/raw/a7fbcb5c0e12d62a448eaa0e260346bf5dcc0feb/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-Next-MLX-4bit/raw/03cba26036330b7553d252c1f3fb899f16cc5ea5/model.safetensors.index.json" }, { "label": "Transformers Qwen3-Next implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/qwen3_next/modeling_qwen3_next.py" } ] }, { "id": "MaziyarPanahi/Qwen3-4B-Instruct-2507-GGUF", "name": "Qwen3-4B-Instruct-2507-GGUF F16", "short_name": "Qwen3 4B Instruct 2507", "license": "unknown", "hf_pipeline_tag": "text-generation", "hf_downloads": 149675, "tags": [ "gguf", "mistral", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:Qwen/Qwen3-4B-Instruct-2507", "base_model:quantized:Qwen/Qwen3-4B-Instruct-2507", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 GGUF (Qwen3ForCausalLM)", "total_params_b": 4.022468096, "active_params_b": 4.022468096, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF F16", "weight_bytes_per_param": 2.0015783185468425, "resident_weight_gb": 8.051284928, "active_weight_gb": 8.045328384, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, pinned model card, stale repo-config check, audited BF16 Qwen3 4B Instruct 2507 base profile, linked-object HEAD checks, and direct GGUF header range read. Legacy fallback values mirror the audited profile: API-selected Qwen3-4B-Instruct-2507.fp16.gguf, exact 8.051284928 GB resident linked file, 8.045328384 GB ordinary text-decode swept tensor traffic, 0.005956544 GB resident-only GGUF overhead, tied token_embd output projection, and FP16 full-context KV. The authoritative profile supersedes the generated mistral/2048-context/2-bit estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-4B-Instruct-2507-GGUF" }, { "label": "Pinned stale repo config", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-4B-Instruct-2507-GGUF/raw/aec29f0e8c31130ba811bec2c774c2ef44888f55/config.json" }, { "label": "BF16 base model card", "url": "https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Qwen3-4B-Instruct-2507-GGUF" }, { "label": "Selected F16 GGUF header audit", "url": "https://huggingface.co/MaziyarPanahi/Qwen3-4B-Instruct-2507-GGUF/resolve/aec29f0e8c31130ba811bec2c774c2ef44888f55/Qwen3-4B-Instruct-2507.fp16.gguf" } ] }, { "id": "Qwen/Qwen3.5-4B-Base", "name": "Qwen3.5-4B-Base F32/BF16", "short_name": "Qwen3.5 4B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 197752, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "conversational", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 4.659865088, "active_params_b": 4.205751296, "layers": 32, "attention_heads": 16, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "F32/BF16", "weight_bytes_per_param": 2.0000016481163843, "resident_weight_gb": 9.319737856, "active_weight_gb": 8.411510272, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, pinned model card/config, Qwen/Qwen3.5-4B config comparison, safetensors index, direct safetensors header range reads, and the Transformers qwen3_5 implementation. Bound weights separate the 9.319737856 GB resident BF16/F32 package from the 8.411510272 GB ordinary text-decode swept subset; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3.5-4B-Base" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3.5-4B-Base" }, { "label": "Pinned model config", "url": "https://huggingface.co/Qwen/Qwen3.5-4B-Base/raw/1001bb4d826a52d1f399e183466143f4da7b741b/config.json" }, { "label": "Pinned instruction config", "url": "https://huggingface.co/Qwen/Qwen3.5-4B/raw/851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Qwen/Qwen3.5-4B-Base/raw/1001bb4d826a52d1f399e183466143f4da7b741b/model.safetensors.index.json" } ] }, { "id": "ggml-org/SmolVLM-500M-Instruct-GGUF", "name": "SmolVLM-500M-Instruct-GGUF GGUF", "short_name": "SmolVLM 500M Instruct", "license": "apache-2.0", "hf_pipeline_tag": "unknown", "hf_downloads": 178205, "tags": [ "gguf", "base_model:HuggingFaceTB/SmolVLM-500M-Instruct", "base_model:quantized:HuggingFaceTB/SmolVLM-500M-Instruct", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "llama (VLlama3ForCausalLM)", "total_params_b": 0.4092528, "active_params_b": 0.361944, "layers": 32, "attention_heads": 15, "kv_heads": 5, "head_dim": 64, "max_context_tokens": 8192, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "F16 GGUF", "weight_bytes_per_param": 2.0046849086921337, "resident_weight_gb": 0.820422912, "active_weight_gb": 0.7240128, "kv_alloc_gb_per_1k_tokens": 0.04096, "kv_read_gb_per_1k_tokens": 0.04096, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, README, HuggingFaceTB base config/API, linked GGUF file sizes, selected main F16 GGUF header range read, and F16 mmproj sidecar header review. The authoritative bounds profile uses the selected SmolVLM-500M-Instruct-f16.gguf artifact: 0.820422912 GB resident linked file, 0.724012800 GB swept text-decode traffic, 0.096410112 GB resident-only token embedding plus GGUF overhead, and FP16 full-context KV. Separate mmproj GGUF files are not included unless explicitly loaded by the workload." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/ggml-org/SmolVLM-500M-Instruct-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/ggml-org/SmolVLM-500M-Instruct-GGUF" }, { "label": "Pinned base model config", "url": "https://huggingface.co/HuggingFaceTB/SmolVLM-500M-Instruct/raw/a7da5b986cb59b408707209984f360a5f4ad7e47/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/HuggingFaceTB/SmolVLM-500M-Instruct" }, { "label": "Selected F16 GGUF header audit", "url": "https://huggingface.co/ggml-org/SmolVLM-500M-Instruct-GGUF/resolve/72e986006ef53e37cdd3f6d4241c90b0f01df376/SmolVLM-500M-Instruct-f16.gguf" } ] }, { "id": "lmstudio-community/Llama-3.2-3B-Instruct-GGUF", "name": "Llama-3.2-3B-Instruct-GGUF GGUF", "short_name": "Llama 3.2 3B Instruct", "license": "llama3.2", "hf_pipeline_tag": "text-generation", "hf_downloads": 251450, "tags": [ "transformers", "gguf", "facebook", "meta", "pytorch", "llama", "llama-3", "text-generation", "en", "de", "fr", "it", "pt", "hi", "es", "th", "base_model:meta-llama/Llama-3.2-3B-Instruct", "base_model:quantized:meta-llama/Llama-3.2-3B-Instruct", "license:llama3.2", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "llama GGUF Q3_K_L (Llama 3.2 3B Instruct)", "total_params_b": 3.212749888, "active_params_b": 3.212749888, "layers": 28, "attention_heads": 24, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q3_K_L", "weight_bytes_per_param": 0.5650447595626841, "resident_weight_gb": 1.815347488, "active_weight_gb": 1.80750976, "kv_alloc_gb_per_1k_tokens": 0.114688, "kv_read_gb_per_1k_tokens": 0.114688, "notes": "Manually audited on 2026-07-06 from current HF API metadata, model card metadata, base-model API metadata, gated base-config access check, linked-object HEAD checks, and a direct selected-file GGUF header range read. The authoritative bounds profile uses the API-selected Llama-3.2-3B-Instruct-Q3_K_L.gguf artifact: 1.815347488 GB resident, 1.807509760 GB swept tensor traffic, 0.007837728 GB resident-only GGUF metadata/header/file overhead, 28 layers, 8 KV heads, 128 key/value head dimension, and 131072-token context." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/Llama-3.2-3B-Instruct-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Llama-3.2-3B-Instruct-GGUF" }, { "label": "Base model card", "url": "https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct" }, { "label": "Base model API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.2-3B-Instruct" }, { "label": "Selected Q3_K_L GGUF header audit", "url": "https://huggingface.co/lmstudio-community/Llama-3.2-3B-Instruct-GGUF/resolve/c91307b5cf18c8106b1f8a6218c26ae4dbfee472/Llama-3.2-3B-Instruct-Q3_K_L.gguf" } ] }, { "id": "lmstudio-community/Qwen3-Coder-Next-MLX-6bit", "name": "Qwen3-Coder-Next-MLX-6bit 6-bit", "short_name": "Qwen3 Coder Next MLX 6bit", "license": "apache-2.0", "hf_pipeline_tag": "unknown", "hf_downloads": 179477, "tags": [ "mlx", "safetensors", "qwen3_next", "base_model:Qwen/Qwen3-Coder-Next", "base_model:quantized:Qwen/Qwen3-Coder-Next", "license:apache-2.0", "6-bit", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_next (Qwen3NextForCausalLM)", "total_params_b": 79.674391296, "active_params_b": 3.580526207, "layers": 48, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "MLX 6-bit affine", "weight_bytes_per_param": 0.8126789750479175, "resident_weight_gb": 64.749702656, "active_weight_gb": 2.909818368, "kv_alloc_gb_per_1k_tokens": 0.024576, "kv_read_gb_per_1k_tokens": 0.024576, "kv_alloc_fixed_gb": 0.077856768, "kv_read_fixed_gb": 0.077856768, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, pinned served MLX config, official BF16 base config/API comparison, safetensors index, direct shard header range reads, and the Transformers qwen3_next implementation. Bound weights use the exact 64.749702656 GB resident MLX package and 2.909818368 GB single-session ordinary text-decode traffic; the audited profile supersedes the generated ideal 6-bit estimate, fixes shared expert count, and adds the fixed DeltaNet runtime state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-Next-MLX-6bit" }, { "label": "Pinned Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3-Coder-Next-MLX-6bit" }, { "label": "Pinned model config", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-Next-MLX-6bit/raw/6b4712e5519c7a7c5612992c5fca1608f42a14c2/config.json" }, { "label": "Pinned BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3-Coder-Next/raw/a7fbcb5c0e12d62a448eaa0e260346bf5dcc0feb/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-Next-MLX-6bit/raw/6b4712e5519c7a7c5612992c5fca1608f42a14c2/model.safetensors.index.json" }, { "label": "Transformers Qwen3-Next implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/qwen3_next/modeling_qwen3_next.py" } ] }, { "id": "RedHatAI/Qwen3.5-9B-FP8-dynamic", "name": "Qwen3.5-9B-FP8-dynamic FP8", "short_name": "Qwen3.5 9B FP8 dynamic", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 178278, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "fp8", "vllm", "llm-compressor", "compressed-tensors", "conversational", "base_model:Qwen/Qwen3.5-9B", "base_model:quantized:Qwen/Qwen3.5-9B", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 9.653104368, "active_params_b": 7.936684544, "layers": 32, "attention_heads": 16, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "hybrid", "weight_precision": "compressed-tensors FP8 dynamic", "weight_bytes_per_param": 1.4510019088193082, "resident_weight_gb": 14.006672864, "active_weight_gb": 10.573833216, "kv_alloc_gb_per_1k_tokens": 0.032768, "kv_read_gb_per_1k_tokens": 0.032768, "kv_alloc_fixed_gb": 0.051904512, "kv_read_fixed_gb": 0.051904512, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, pinned compressed-tensors config and recipe, Qwen3.5 9B base config comparison, direct safetensors shard-header range reads, and the Transformers qwen3_5 implementation. The authoritative bounds profile uses exact stored bytes: 14.006672864 GB resident, 10.573833216 GB ordinary text-decode swept traffic, 3.432839648 GB resident-only input embedding/vision/MTP tensors, 8 full-attention layers, and a fixed DeltaNet recurrent-state charge of 0.051904512 GB per session/token." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/RedHatAI/Qwen3.5-9B-FP8-dynamic" }, { "label": "Pinned Hugging Face API metadata", "url": "https://huggingface.co/api/models/RedHatAI/Qwen3.5-9B-FP8-dynamic" }, { "label": "Pinned model config", "url": "https://huggingface.co/RedHatAI/Qwen3.5-9B-FP8-dynamic/raw/790f0576d2d77dd5322aa0603a470bd9e3a3d1f6/config.json" }, { "label": "Pinned quantization recipe", "url": "https://huggingface.co/RedHatAI/Qwen3.5-9B-FP8-dynamic/raw/790f0576d2d77dd5322aa0603a470bd9e3a3d1f6/recipe.yaml" }, { "label": "Pinned BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3.5-9B/raw/c202236235762e1c871ad0ccb60c8ee5ba337b9a/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/RedHatAI/Qwen3.5-9B-FP8-dynamic/raw/790f0576d2d77dd5322aa0603a470bd9e3a3d1f6/model.safetensors.index.json" }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py" } ] }, { "id": "google/gemma-1.1-2b-it", "name": "gemma-1.1-2b-it BF16", "short_name": "gemma 1.1 2b it", "license": "gemma", "hf_pipeline_tag": "text-generation", "hf_downloads": 198323, "tags": [ "transformers", "safetensors", "gemma", "text-generation", "conversational", "arxiv:2312.11805", "arxiv:2009.03300", "arxiv:1905.07830", "arxiv:1911.11641", "arxiv:1904.09728", "arxiv:1905.10044", "arxiv:1907.10641", "arxiv:1811.00937", "arxiv:1809.02789", "arxiv:1911.01547", "arxiv:1705.03551", "arxiv:2107.03374", "arxiv:2108.07732", "arxiv:2110.14168", "arxiv:2304.06364", "arxiv:2206.04615", "arxiv:1804.06876", "arxiv:2110.08193", "license:gemma", "eval-results", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gated config unavailable", "total_params_b": 2.506172416, "active_params_b": 2.506172416, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 5.012344832, "active_weight_gb": 5.012344832, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manually checked on 2026-07-06 from current Hugging Face API metadata and raw/HF CLI gated-access checks. The API exposes GemmaForCausalLM/model_type gemma and BF16 safetensors total 2506172416, but config and tensor headers are gated from the configured osolmaz token, so context, KV geometry, tied embeddings, and swept decode traffic are not verified. The authoritative profile fails closed with unsupported-gated-config." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-1.1-2b-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-1.1-2b-it" }, { "label": "Pinned gated config access check", "url": "https://huggingface.co/google/gemma-1.1-2b-it/raw/d750f5eceb83e978c09e2b3597c2a8784e381022/config.json" } ] }, { "id": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4", "name": "Meta-Llama-3.1-8B-Instruct-AWQ-INT4 4-bit", "short_name": "Meta Llama 3.1 8B Instruct AWQ INT4", "license": "llama3.1", "hf_pipeline_tag": "text-generation", "hf_downloads": 194297, "tags": [ "transformers", "safetensors", "llama", "text-generation", "llama-3.1", "meta", "autoawq", "conversational", "en", "de", "fr", "it", "pt", "hi", "es", "th", "license:llama3.1", "text-generation-inference", "endpoints_compatible", "4-bit", "awq", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "llama (LlamaForCausalLM)", "total_params_b": 8.030261248, "active_params_b": 7.504924672, "layers": 32, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "AWQ INT4", "weight_bytes_per_param": 0.7132837170679307, "resident_weight_gb": 5.727854592, "active_weight_gb": 4.67718144, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, model card, pinned AWQ config, existing audited Llama 3.1 quantized config comparison, safetensors index, linked-object HEAD checks, and direct shard-header range reads. The authoritative bounds profile uses exact stored bytes: 5.727854592 GB resident, 4.677181440 GB ordinary text-decode swept traffic, 1.050673152 GB resident-only input embedding, and full-context FP16 KV." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4" }, { "label": "Pinned model config", "url": "https://huggingface.co/hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4/raw/db1f81ad4b8c7e39777509fac66c652eb0a52f91/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4/raw/db1f81ad4b8c7e39777509fac66c652eb0a52f91/model.safetensors.index.json" }, { "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" } ] }, { "id": "bullpoint/Qwen3-Coder-Next-AWQ-4bit", "name": "Qwen3-Coder-Next-AWQ-4bit 4-bit", "short_name": "Qwen3 Coder Next AWQ 4bit", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 154238, "tags": [ "transformers", "safetensors", "qwen3_next", "text-generation", "qwen", "qwen3", "qwen3-coder", "qwen3-next", "moe", "mixture-of-experts", "awq", "4bit", "int4", "quantized", "llm-compressor", "compressed-tensors", "code", "coding", "coder", "agent", "agentic", "tool-use", "function-calling", "vllm", "sglang", "conversational", "en", "zh", "dataset:nvidia/Llama-Nemotron-Post-Training-Dataset", "base_model:Qwen/Qwen3-Coder-Next", "base_model:quantized:Qwen/Qwen3-Coder-Next", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_next (Qwen3NextForCausalLM)", "total_params_b": 79.674391296, "active_params_b": 3.56376448, "layers": 48, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "compressed-tensors AWQ INT4", "weight_bytes_per_param": 0.6051842087737511, "resident_weight_gb": 48.217683456, "active_weight_gb": 4.95699968, "kv_alloc_gb_per_1k_tokens": 0.024576, "kv_read_gb_per_1k_tokens": 0.024576, "kv_alloc_fixed_gb": 0.077856768, "kv_read_fixed_gb": 0.077856768, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, pinned compressed-tensors AWQ config and recipe, official BF16 base config comparison, safetensors index, direct shard-header range reads, and the Transformers qwen3_next implementation. Bound weights use the exact 48.217683456 GB resident package and 4.956999680 GB single-session ordinary text-decode traffic; the audited profile supersedes the generated ideal 4-bit estimate, fixes shared expert count, and adds the fixed DeltaNet runtime state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/bullpoint/Qwen3-Coder-Next-AWQ-4bit" }, { "label": "Pinned Hugging Face API metadata", "url": "https://huggingface.co/api/models/bullpoint/Qwen3-Coder-Next-AWQ-4bit" }, { "label": "Pinned model config", "url": "https://huggingface.co/bullpoint/Qwen3-Coder-Next-AWQ-4bit/raw/b035ac132045fd726eaa1ce53888df39cbb30291/config.json" }, { "label": "Pinned compression recipe", "url": "https://huggingface.co/bullpoint/Qwen3-Coder-Next-AWQ-4bit/raw/b035ac132045fd726eaa1ce53888df39cbb30291/recipe.yaml" }, { "label": "Pinned BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3-Coder-Next/raw/a7fbcb5c0e12d62a448eaa0e260346bf5dcc0feb/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/bullpoint/Qwen3-Coder-Next-AWQ-4bit/resolve/b035ac132045fd726eaa1ce53888df39cbb30291/model.safetensors.index.json" }, { "label": "Transformers Qwen3-Next implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/qwen3_next/modeling_qwen3_next.py" } ] }, { "id": "Mungert/Hunyuan-MT-7B-GGUF", "name": "Hunyuan-MT-7B-GGUF GGUF", "short_name": "Hunyuan MT 7B", "license": "unknown", "hf_pipeline_tag": "translation", "hf_downloads": 170812, "tags": [ "transformers", "gguf", "translation", "zh", "en", "fr", "pt", "es", "ja", "tr", "ru", "ar", "ko", "th", "it", "de", "vi", "ms", "id", "tl", "hi", "pl", "cs", "nl", "km", "my", "fa", "gu", "ur", "te", "mr", "he", "bn", "ta", "uk", "bo", "kk", "mn", "ug", "arxiv:2509.05209", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "hunyuan-dense (HunYuanDenseV1ForCausalLM GGUF)", "total_params_b": 7.504932864, "active_params_b": 7.504932864, "layers": 32, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF BF16/Q8_0", "weight_bytes_per_param": 1.447791112978569, "resident_weight_gb": 10.865575104, "active_weight_gb": 10.858053632, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, pinned Mungert model card, Tencent base config, linked-object HEAD checks, and direct GGUF header range reads. Bound weights use the exact 10.865575104 GB API-selected BF16/Q8_0 linked artifact and 10.858053632 GB tensor traffic; token_embd.weight is swept because this tied-embedding model has no separate output.weight tensor." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Mungert/Hunyuan-MT-7B-GGUF" }, { "label": "Pinned Hugging Face API metadata", "url": "https://huggingface.co/api/models/Mungert/Hunyuan-MT-7B-GGUF" }, { "label": "Pinned Tencent base config", "url": "https://huggingface.co/tencent/Hunyuan-MT-7B/raw/9305c78383f0bcc94358e08667ee2c76107877e3/config.json" }, { "label": "Selected GGUF linked artifact and header audit", "url": "https://huggingface.co/Mungert/Hunyuan-MT-7B-GGUF/resolve/61e98ae605cc4fe9581fd3ff1052a271843e4d64/Hunyuan-MT-7B-bf16_q8_0.gguf" } ] }, { "id": "unsloth/GLM-5.2-GGUF", "name": "GLM-5.2 BF16 GGUF", "short_name": "GLM 5.2", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 392857, "tags": [ "gguf", "glm_moe_dsa", "unsloth", "text-generation", "en", "zh", "arxiv:2602.15763", "arxiv:2603.12201", "base_model:zai-org/GLM-5.2", "base_model:quantized:zai-org/GLM-5.2", "license:mit", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "MoE transformer", "detail": "glm-dsa BF16 GGUF (GlmMoeDsaForCausalLM via llama.cpp)", "total_params_b": 753.864139008, "active_params_b": 40.1021376, "layers": 78, "attention_heads": 64, "kv_heads": 1, "head_dim": 576, "max_context_tokens": 1048576, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "GGUF BF16", "weight_bytes_per_param": 2.000344551463002, "resident_weight_gb": 1507.988023008, "active_weight_gb": 80.442571776, "kv_alloc_gb_per_1k_tokens": 0.089856, "kv_read_gb_per_1k_tokens": 0.089856, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, pinned Unsloth card, pinned Z.ai base config, direct GGUF split header range reads, and current llama.cpp GLM_DSA/cache source review. Bound weights use the exact 1507.988023008 GB API-selected BF16 split linked artifact. Active single-session traffic charges current llama.cpp ordinary decode: 35.144088576 GB fixed used tensors plus 8 routed experts at 5.6623104 GB each. Resident-only bytes include token embeddings, layer-78 MTP/NextN tensors, GGUF overhead, and optional DSA indexer tensors that current llama.cpp loads as not-required but does not execute." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/GLM-5.2-GGUF" }, { "label": "Pinned Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/GLM-5.2-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/zai-org/GLM-5.2/raw/b4734de4facf877f85769a911abafc5283eab3d9/config.json" }, { "label": "Selected BF16 GGUF split and header audit", "url": "https://huggingface.co/unsloth/GLM-5.2-GGUF/tree/abc55e72527792c6e77069c99b4cb7de16fa9f23/BF16" }, { "label": "llama.cpp GLM_DSA source review", "url": "https://github.com/ggml-org/llama.cpp/blob/cb295bf59663cd3577389315636772f4060bd1f5/src/models/glm-dsa.cpp" } ] }, { "id": "sugoitoolkit/Sugoi-14B-Ultra-GGUF", "name": "Sugoi 14B Ultra F16 GGUF", "short_name": "Sugoi 14B Ultra", "license": "apache-2.0", "hf_pipeline_tag": "translation", "hf_downloads": 187132, "tags": [ "gguf", "translation", "ja", "en", "base_model:sugoitoolkit/Sugoi-14B-Ultra-HF", "base_model:quantized:sugoitoolkit/Sugoi-14B-Ultra-HF", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 14.770033664, "active_params_b": 13.991465984, "layers": 48, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF F16", "weight_bytes_per_param": 2.000517839848845, "resident_weight_gb": 29.54771584, "active_weight_gb": 27.984613376, "kv_alloc_gb_per_1k_tokens": 0.196608, "kv_read_gb_per_1k_tokens": 0.196608, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, pinned Sugoi card, pinned Sugoi HF base config, linked-object HEAD checks, and a direct selected GGUF header range read. Bound weights use the exact 29.547715840 GB API-selected Sugoi-14B-Ultra-F16.gguf linked artifact and 27.984613376 GB swept tensor traffic; token_embd.weight is resident-only because the GGUF stores a separate output.weight tensor." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/sugoitoolkit/Sugoi-14B-Ultra-GGUF" }, { "label": "Pinned Hugging Face API metadata", "url": "https://huggingface.co/api/models/sugoitoolkit/Sugoi-14B-Ultra-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/sugoitoolkit/Sugoi-14B-Ultra-HF/raw/9e71532aff2389b5e00f00eab72ee59f77de17a4/config.json" }, { "label": "Selected GGUF linked artifact and header audit", "url": "https://huggingface.co/sugoitoolkit/Sugoi-14B-Ultra-GGUF/resolve/d8dd836bf519a604530779cbf5ce13ffb35c3eeb/Sugoi-14B-Ultra-F16.gguf" } ] }, { "id": "Jackrong/Qwopus3.6-35B-A3B-v1-GGUF", "name": "Qwopus3.6 35B A3B v1 IQ4_XS GGUF", "short_name": "Qwopus3.6 35B A3B v1", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 178977, "tags": [ "transformers", "gguf", "llama.cpp", "image-text-to-text", "vision", "multimodal", "text-generation-inference", "unsloth", "conversational", "qwen3_6", "moe", "reasoning", "chain-of-thought", "lora", "sft", "tool-use", "function-calling", "long-context", "en", "zh", "es", "ru", "ja", "dataset:Jackrong/Claude-opus-4.7-TraceInversion-5000x", "dataset:Jackrong/Claude-opus-4.6-TraceInversion-9000x", "base_model:unsloth/Qwen3.6-35B-A3B", "base_model:adapter:unsloth/Qwen3.6-35B-A3B", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen35moe / qwen3_5_moe_text", "total_params_b": 34.660610688, "active_params_b": 2.946429568, "layers": 40, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "GGUF IQ4_XS", "weight_bytes_per_param": 0.5464217778066173, "resident_weight_gb": 18.939312512, "active_weight_gb": 1.877002752, "kv_alloc_gb_per_1k_tokens": 0.02048, "kv_read_gb_per_1k_tokens": 0.02048, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, pinned Unsloth Qwen3.6 config, selected linked-object HEAD checks, and a direct GGUF header range read. Live API downloads were 97662 during audit, below the 100k catalog threshold, so this row keeps the generated catalog snapshot count 178977. The selected API artifact is Qwopus3.6-35B-A3B-v1-IQ4_XS.gguf at 18.939312512 GB resident; ordinary single-session swept traffic is 1.877002752 GB before KV/state reads. The audited profile adds the fixed DeltaNet state charge that these legacy per-1K KV fields cannot fully express." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Jackrong/Qwopus3.6-35B-A3B-v1-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Jackrong/Qwopus3.6-35B-A3B-v1-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/unsloth/Qwen3.6-35B-A3B/raw/2ab40a9acc6d567889ca4d4e59feb2da56121454/config.json" }, { "label": "Selected GGUF header audit", "url": "https://huggingface.co/Jackrong/Qwopus3.6-35B-A3B-v1-GGUF/resolve/bd297a4fe8dd8f9987c2c7a5b1029834c9a76c3f/Qwopus3.6-35B-A3B-v1-IQ4_XS.gguf" } ] }, { "id": "nvidia/Llama-3_3-Nemotron-Super-49B-v1", "name": "Llama-3_3-Nemotron-Super-49B-v1 BF16", "short_name": "Llama 3 3 Nemotron Super 49B v1", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 188263, "tags": [ "transformers", "safetensors", "nemotron-nas", "text-generation", "nvidia", "llama-3", "pytorch", "conversational", "custom_code", "en", "arxiv:2411.19146", "arxiv:2505.00949", "arxiv:2502.00203", "license:other", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "nemotron-nas (DeciLMForCausalLM)", "total_params_b": 49.867145216, "active_params_b": 49.867145216, "layers": 80, "attention_heads": 64, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 99.734290432, "active_weight_gb": 97.632944128, "kv_alloc_gb_per_1k_tokens": 0.200704, "kv_read_gb_per_1k_tokens": 0.200704, "notes": "Manually audited on 2026-07-06 from current Hugging Face API metadata, pinned served DeciLM config, custom runtime code, block config dataclasses, safetensors index, and direct safetensors shard header range reads. downloads=188263. Exact BF16 tensor payload bytes: 99.734290432 GB resident, 97.632944128 GB swept ordinary text traffic, and 2.101346304 GB resident-only input embedding. The audited DeciLM profile supersedes the generated estimate, which incorrectly charged 64 KV heads across all 80 layers." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1" }, { "label": "Model config", "url": "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1/raw/387156d8d6868c19f3472fa607aa9bfc4f662333/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/Llama-3_3-Nemotron-Super-49B-v1" }, { "label": "Custom DeciLM runtime", "url": "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1/raw/387156d8d6868c19f3472fa607aa9bfc4f662333/modeling_decilm.py" }, { "label": "Safetensors index", "url": "https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1/raw/387156d8d6868c19f3472fa607aa9bfc4f662333/model.safetensors.index.json" } ] }, { "id": "google/gemma-4-E4B-it-qat-w4a16-ct", "name": "gemma-4-E4B-it-qat-w4a16-ct 4-bit", "short_name": "gemma 4 E4B it qat w4a16 ct", "license": "apache-2.0", "hf_pipeline_tag": "any-to-any", "hf_downloads": 291658, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "any-to-any", "base_model:google/gemma-4-E4B-it-qat-q4_0-unquantized", "base_model:quantized:google/gemma-4-E4B-it-qat-q4_0-unquantized", "license:apache-2.0", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 8.736340888, "active_params_b": 4.768591576, "layers": 42, "attention_heads": 8, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "W4A16 compressed-tensors", "weight_bytes_per_param": 1.3178855495227555, "resident_weight_gb": 11.513497412, "active_weight_gb": 3.577998788, "kv_alloc_gb_per_1k_tokens": 0.03735552, "kv_read_gb_per_1k_tokens": 0.06537216, "notes": "Audited profile-backed row from live Hugging Face metadata on 2026-07-06; downloads=291658; data quality: API metadata, model card, served compressed-tensors config, QAT source config comparison, upstream Transformers Gemma 4 implementation review, and direct safetensors header grouping. Resident weight uses exact stored tensor bytes; active weight excludes resident-only input embedding, PLE lookup, audio, and vision tensors. KV allocation/read summaries reflect Gemma 4 E4B shared-KV behavior at a 1k-token reference point; production bounds use the audited profile formulas." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-4-E4B-it-qat-w4a16-ct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-E4B-it-qat-w4a16-ct" }, { "label": "Pinned model config", "url": "https://huggingface.co/google/gemma-4-E4B-it-qat-w4a16-ct/raw/ef0a4c43726bde42a3ca04fd300397c0b8b3c3f0/config.json" }, { "label": "Pinned safetensors artifact", "url": "https://huggingface.co/google/gemma-4-E4B-it-qat-w4a16-ct/resolve/ef0a4c43726bde42a3ca04fd300397c0b8b3c3f0/model.safetensors" }, { "label": "QAT source config", "url": "https://huggingface.co/google/gemma-4-E4B-it-qat-q4_0-unquantized/raw/dfc5b925ddb1d41aaf1fe9679abdcfb0805e1aa6/config.json" } ] }, { "id": "txn545/Qwen3.5-122B-A10B-NVFP4", "name": "Qwen3.5-122B-A10B-NVFP4 NVFP4", "short_name": "Qwen3.5 122B A10B NVFP4", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 129110, "tags": [ "Model Optimizer", "safetensors", "ModelOpt", "Qwen3.5", "quantized", "FP4", "fp4", "text-generation", "base_model:Qwen/Qwen3.5-122B-A10B", "base_model:finetune:Qwen/Qwen3.5-122B-A10B", "license:apache-2.0", "8-bit", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_5_moe_text", "total_params_b": 125.086497008, "active_params_b": 10, "layers": 48, "attention_heads": 32, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "ModelOpt NVFP4/FP8/BF16/F32", "weight_bytes_per_param": 0.5, "resident_weight_gb": 82.823706304, "active_weight_gb": 12.156418272, "kv_alloc_gb_per_1k_tokens": 0.012288, "kv_read_gb_per_1k_tokens": 0.012288, "kv_alloc_fixed_gb": 0.154533888, "kv_read_fixed_gb": 0.154533888, "expert_param_b": 0.50961024, "fixed_param_b": 20.235954624, "shared_expert_param_b": 0, "notes": "Manually corrected on 2026-07-06 from current Hugging Face API metadata, pinned ModelOpt NVFP4 config, hf_quant_config, BF16 base config comparison, safetensors index, and direct range-read safetensors shard headers. Legacy fallback values mirror the audited profile: exact stored resident bytes, fixed always-on hybrid/shared-expert traffic, expected distinct routed experts, FP8 attention KV, and fixed BF16-convolution/F32-SSM DeltaNet recurrent state. The authoritative profile supersedes the generated 62.5433 GB resident estimate, the missing shared-expert count, and the all-layers full-context KV estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/txn545/Qwen3.5-122B-A10B-NVFP4" }, { "label": "Model config", "url": "https://huggingface.co/txn545/Qwen3.5-122B-A10B-NVFP4/raw/cdd3d924f4c96cf9ce429ad34df90b8d7654f4b9/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/txn545/Qwen3.5-122B-A10B-NVFP4" }, { "label": "ModelOpt quantization config", "url": "https://huggingface.co/txn545/Qwen3.5-122B-A10B-NVFP4/raw/cdd3d924f4c96cf9ce429ad34df90b8d7654f4b9/hf_quant_config.json" }, { "label": "Base config", "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B/raw/dc4d348443bc740c68e2d77492492c11606384d5/config.json" }, { "label": "Safetensors index", "url": "https://huggingface.co/txn545/Qwen3.5-122B-A10B-NVFP4/resolve/cdd3d924f4c96cf9ce429ad34df90b8d7654f4b9/model.safetensors.index.json" } ] }, { "id": "LilaRest/gemma-4-31B-it-NVFP4-turbo", "name": "gemma-4-31B-it-NVFP4-turbo NVFP4", "short_name": "gemma 4 31B it NVFP4 turbo", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 182086, "tags": [ "transformers", "safetensors", "gemma4", "text-generation", "gemma-4-31b-it", "nvfp4", "modelopt", "vllm", "quantized", "nvidia", "lighthouse", "conversational", "base_model:google/gemma-4-31B-it", "base_model:quantized:google/gemma-4-31B-it", "license:apache-2.0", "model-index", "endpoints_compatible", "4-bit", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text (Gemma4ForCausalLM text-only)", "total_params_b": 32.52776664, "active_params_b": 32.52776664, "layers": 60, "attention_heads": 32, "kv_heads": 16, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "NVFP4 ModelOpt + BF16/F8/U8 side tensors", "weight_bytes_per_param": 0.5, "resident_weight_gb": 19.29502292, "active_weight_gb": 19.29502292, "kv_alloc_gb_per_1k_tokens": 0.41984, "kv_read_gb_per_1k_tokens": 0.41984, "notes": "Manually corrected on 2026-07-06 from current Hugging Face API metadata, pinned ModelOpt config, Google base config comparison, safetensors index, and direct range-read safetensors shard headers. The authoritative profile supersedes the generated half-byte resident estimate and all-layers full-context KV estimate: this is a text-only Gemma4ForCausalLM package with exact 19.295022920 GB stored resident/swept bytes and layered FP8 KV over 10 full-context layers plus 50 sliding-window layers." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/LilaRest/gemma-4-31B-it-NVFP4-turbo" }, { "label": "Model config", "url": "https://huggingface.co/LilaRest/gemma-4-31B-it-NVFP4-turbo/raw/77fb4a2380f1d167473fb11fa99e3d982bebbc2b/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-31B-it" }, { "label": "Base model config", "url": "https://huggingface.co/google/gemma-4-31B-it/raw/3548789868c5356dbf307c98e6f609007b82b3eb/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/LilaRest/gemma-4-31B-it-NVFP4-turbo" }, { "label": "Safetensors index", "url": "https://huggingface.co/LilaRest/gemma-4-31B-it-NVFP4-turbo/resolve/77fb4a2380f1d167473fb11fa99e3d982bebbc2b/model.safetensors.index.json" } ] }, { "id": "MaziyarPanahi/Mistral-7B-Instruct-v0.3-GGUF", "name": "Mistral-7B-Instruct-v0.3-GGUF FP16", "short_name": "Mistral 7B Instruct v0.3", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 142761, "tags": [ "transformers", "gguf", "mistral", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "safetensors", "text-generation", "conversational", "license:apache-2.0", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "region:us", "base_model:mistralai/Mistral-7B-Instruct-v0.3", "base_model:quantized:mistralai/Mistral-7B-Instruct-v0.3" ], "architecture": { "type": "Dense transformer", "detail": "mistral (GGUF FP16)", "total_params_b": 7.248023552, "active_params_b": 7.113805824, "layers": 32, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF FP16/F32", "weight_bytes_per_param": 2, "resident_weight_gb": 14.497337312, "active_weight_gb": 14.228144128, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "notes": "Manually corrected on 2026-07-06 from current Hugging Face API metadata, pinned Mistral base config, selected GGUF linked-object HEAD checks, and direct range-read GGUF header/tensor index. The HF API gguf.totalFileSize selects Mistral-7B-Instruct-v0.3.fp16.gguf, so this row intentionally models the FP16 artifact rather than lower-bit GGUF siblings advertised by tags. Exact selected file residency is 14.497337312 GB, ordinary swept tensor traffic is 14.228144128 GB after excluding the untied input embedding and GGUF overhead, and KV uses full-context FP16 K/V over 32 layers." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/Mistral-7B-Instruct-v0.3-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/MaziyarPanahi/Mistral-7B-Instruct-v0.3-GGUF/raw/ce89f595755a4bf2e2e05d155cc43cb847c78978/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3" }, { "label": "Base model config", "url": "https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3/raw/c170c708c41dac9275d15a8fff4eca08d52bab71/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Mistral-7B-Instruct-v0.3-GGUF" }, { "label": "Selected FP16 GGUF", "url": "https://huggingface.co/MaziyarPanahi/Mistral-7B-Instruct-v0.3-GGUF/resolve/ce89f595755a4bf2e2e05d155cc43cb847c78978/Mistral-7B-Instruct-v0.3.fp16.gguf" } ] }, { "id": "MaziyarPanahi/gemma-3-4b-it-GGUF", "name": "gemma-3-4b-it-GGUF FP16", "short_name": "Gemma 3 4B IT", "license": "gemma", "hf_pipeline_tag": "text-generation", "hf_downloads": 142375, "tags": [ "gguf", "mistral", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:google/gemma-3-4b-it", "base_model:quantized:google/gemma-3-4b-it", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "gemma3 (GGUF FP16)", "total_params_b": 3.880099328, "active_params_b": 3.880099328, "layers": 34, "attention_heads": 8, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF FP16/F32", "weight_bytes_per_param": 2, "resident_weight_gb": 7.767474368, "active_weight_gb": 7.760934912, "kv_alloc_gb_per_1k_tokens": 0.02169634816, "kv_read_gb_per_1k_tokens": 0.024281088, "notes": "Manually corrected on 2026-07-06 from current Hugging Face API metadata, selected GGUF linked-object HEAD checks, Google Gemma 3 local/global attention documentation, and direct range-read GGUF header/tensor index. The HF API gguf.totalFileSize selects gemma-3-4b-it.fp16.gguf, so this row intentionally models the FP16 artifact rather than lower-bit GGUF siblings advertised by tags. Exact selected file residency is 7.767474368 GB and ordinary swept tensor traffic is 7.760934912 GB. KV uses Gemma 3 layered FP16 K/V over five full-context layers plus 29 sliding-window layers; the legacy catalog coefficients are effective display coefficients for the default compare contexts, while the profile contains the authoritative nonlinear layered KV adapter." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/gemma-3-4b-it-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/MaziyarPanahi/gemma-3-4b-it-GGUF/raw/55ebd63de390fd82075d7ec5d1cf20486292d150/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-3-4b-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/gemma-3-4b-it-GGUF" }, { "label": "Selected FP16 GGUF", "url": "https://huggingface.co/MaziyarPanahi/gemma-3-4b-it-GGUF/resolve/55ebd63de390fd82075d7ec5d1cf20486292d150/gemma-3-4b-it.fp16.gguf" } ] }, { "id": "MaziyarPanahi/Meta-Llama-3-8B-Instruct-GGUF", "name": "Meta-Llama-3-8B-Instruct-GGUF FP16", "short_name": "Meta Llama 3 8B Instruct", "license": "llama3", "hf_pipeline_tag": "text-generation", "hf_downloads": 142135, "tags": [ "transformers", "gguf", "mistral", "facebook", "meta", "pytorch", "llama", "llama-3", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "16-bit", "GGUF", "text-generation", "en", "base_model:meta-llama/Meta-Llama-3-8B-Instruct", "base_model:quantized:meta-llama/Meta-Llama-3-8B-Instruct", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "llama (GGUF FP16)", "total_params_b": 8.030261248, "active_params_b": 7.504924672, "layers": 32, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 8192, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF FP16/F32", "weight_bytes_per_param": 2, "resident_weight_gb": 16.06940384, "active_weight_gb": 15.010381824, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "notes": "Manually corrected on 2026-07-06 from current Hugging Face API metadata, selected GGUF linked-object HEAD checks, the bundled Meta Llama 3 card text, existing gated-base access profile, and direct range-read GGUF header/tensor index. The HF API gguf.totalFileSize selects Meta-Llama-3-8B-Instruct.fp16.gguf, so this row intentionally models the FP16 artifact rather than lower-bit GGUF siblings advertised by tags. Exact selected file residency is 16.069403840 GB, ordinary swept tensor traffic is 15.010381824 GB after excluding the untied input embedding and GGUF overhead, and KV uses full-context FP16 K/V over 32 layers." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/Meta-Llama-3-8B-Instruct-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/MaziyarPanahi/Meta-Llama-3-8B-Instruct-GGUF/raw/4ece958b356f2ec00338e5898ba0a7254d976baf/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Meta-Llama-3-8B-Instruct-GGUF" }, { "label": "Selected FP16 GGUF", "url": "https://huggingface.co/MaziyarPanahi/Meta-Llama-3-8B-Instruct-GGUF/resolve/4ece958b356f2ec00338e5898ba0a7254d976baf/Meta-Llama-3-8B-Instruct.fp16.gguf" } ] }, { "id": "meta-llama/Meta-Llama-3-70B", "name": "Meta-Llama-3-70B BF16", "short_name": "Meta Llama 3 70B", "license": "llama3", "hf_pipeline_tag": "text-generation", "hf_downloads": 149077, "tags": [ "transformers", "safetensors", "llama", "text-generation", "facebook", "meta", "pytorch", "llama-3", "en", "license:llama3", "eval-results", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gated config unavailable", "total_params_b": 70.553706496, "active_params_b": 70.553706496, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 141.107412992, "active_weight_gb": 141.107412992, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata and access checks. The API reports BF16 safetensors parameters, but raw config, README, and safetensors index requests returned HTTP 401 because the repository requires approval. Production bounds use the fail-closed unsupported profile until config and tensor evidence are directly accessible." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/meta-llama/Meta-Llama-3-70B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/meta-llama/Meta-Llama-3-70B" }, { "label": "Gated config access check", "url": "https://huggingface.co/meta-llama/Meta-Llama-3-70B/raw/c82494877ce7f6d7d317c56ec081328e382c72fe/config.json" } ] }, { "id": "CohereLabs/aya-vision-8b", "name": "aya-vision-8b F16", "short_name": "aya vision 8b", "license": "cc-by-nc-4.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 161292, "tags": [ "transformers", "safetensors", "aya_vision", "image-text-to-text", "conversational", "en", "fr", "de", "es", "it", "pt", "ja", "ko", "zh", "ar", "el", "fa", "pl", "id", "cs", "he", "hi", "nl", "ro", "ru", "tr", "uk", "vi", "arxiv:2412.04261", "license:cc-by-nc-4.0", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gated config unavailable", "total_params_b": 8.631842032, "active_params_b": 8.631842032, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "F16", "weight_bytes_per_param": 2, "resident_weight_gb": 17.263684064, "active_weight_gb": 17.263684064, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata and access checks. The API reports F16 safetensors parameters, but raw config, README, and safetensors index requests returned HTTP 401 because the repository requires approval. Production bounds use the fail-closed unsupported profile until config and tensor evidence are directly accessible." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/CohereLabs/aya-vision-8b" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/CohereLabs/aya-vision-8b" }, { "label": "Gated config access check", "url": "https://huggingface.co/CohereLabs/aya-vision-8b/raw/d7601ac8fd3baf6f59c5fc9514c618ef0f24268f/config.json" } ] }, { "id": "cyankiwi/gemma-4-26B-A4B-it-AWQ-8bit", "name": "gemma-4-26B-A4B-it-AWQ-8bit 8-bit", "short_name": "gemma 4 26B A4B it AWQ 8bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 169666, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "conversational", "base_model:google/gemma-4-26B-A4B-it", "base_model:quantized:google/gemma-4-26B-A4B-it", "license:apache-2.0", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "gemma4_text", "total_params_b": 26.554316366, "active_params_b": 4, "layers": 30, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "compressed-tensors AWQ INT8", "weight_bytes_per_param": 1, "resident_weight_gb": 29.160467612, "active_weight_gb": 5.26610438, "kv_alloc_gb_per_1k_tokens": 0.20992, "kv_read_gb_per_1k_tokens": 0.20992, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, served AWQ config, audited 4-bit sibling comparison, safetensors index, and direct range-read shard headers. Bound weights separate the exact 29.160467612 GB resident AWQ package from 3.749519420 GB fixed ordinary text traffic plus 0.189573120 GB per routed expert index; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/cyankiwi/gemma-4-26B-A4B-it-AWQ-8bit" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/cyankiwi/gemma-4-26B-A4B-it-AWQ-8bit" }, { "label": "Pinned model config", "url": "https://huggingface.co/cyankiwi/gemma-4-26B-A4B-it-AWQ-8bit/raw/7d086af42fd9ee78a877a1206b28dedd4f99e5e6/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/cyankiwi/gemma-4-26B-A4B-it-AWQ-8bit/raw/7d086af42fd9ee78a877a1206b28dedd4f99e5e6/model.safetensors.index.json" } ] }, { "id": "zai-org/GLM-4.5", "name": "GLM-4.5 BF16/F32", "short_name": "GLM 4.5", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 144724, "tags": [ "transformers", "safetensors", "glm4_moe", "text-generation", "conversational", "en", "zh", "arxiv:2508.06471", "license:mit", "eval-results", "endpoints_compatible", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "glm4_moe (Glm4MoeForCausalLM)", "total_params_b": 358.337791296, "active_params_b": 32.856334144, "layers": 92, "attention_heads": 96, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": 160, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "BF16/F32", "weight_bytes_per_param": 2.0000000803710933, "resident_weight_gb": 716.675611392, "active_weight_gb": 65.712668288, "kv_alloc_gb_per_1k_tokens": 0.376832, "kv_read_gb_per_1k_tokens": 0.376832, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, pinned served config, generation config, safetensors index, direct range-read shard headers, and the Transformers glm4_moe implementation. Bound weights separate the exact 716.675611392 GB resident BF16/F32 package from 65.712668288 GB ordinary text-decode swept traffic: 32.116293248 GB fixed traffic plus 8 routed expert groups at 4.199546880 GB each. The input embedding and model.layers.92 next-token-prediction tensor package are resident-only for ordinary decode." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/zai-org/GLM-4.5" }, { "label": "Pinned model config", "url": "https://huggingface.co/zai-org/GLM-4.5/raw/cbb2c7cfb52fa128a9660cb1a7a78e017899e115/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/zai-org/GLM-4.5" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/zai-org/GLM-4.5/raw/cbb2c7cfb52fa128a9660cb1a7a78e017899e115/model.safetensors.index.json" } ] }, { "id": "MaziyarPanahi/Meta-Llama-3.1-8B-Instruct-GGUF", "name": "Meta-Llama-3.1-8B-Instruct-GGUF GGUF Q2_K", "short_name": "Meta Llama 3.1 8B Instruct", "license": "llama3.1", "hf_pipeline_tag": "text-generation", "hf_downloads": 136669, "tags": [ "gguf", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "en", "de", "fr", "it", "pt", "hi", "es", "th", "arxiv:2204.05149", "base_model:meta-llama/Llama-3.1-8B-Instruct", "base_model:quantized:meta-llama/Llama-3.1-8B-Instruct", "license:llama3.1", "region:us", "imatrix", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "llama GGUF", "total_params_b": 8.030261248, "active_params_b": 7.504924672, "layers": 32, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q2_K", "weight_bytes_per_param": 0.3958939453920989, "resident_weight_gb": 3.179131808, "active_weight_gb": 2.998919168, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "notes": "Manual correction on 2026-07-06 from current Hugging Face API metadata, pinned model card, gated base-profile checks, linked-object HEAD checks, missing repo-config check, and direct range-read GGUF header/tensor index. The API-selected artifact is Meta-Llama-3.1-8B-Instruct.Q2_K.gguf: exact 3.179131808 GB resident linked file, 2.998919168 GB ordinary text-decode swept tensor traffic, 0.180212640 GB resident-only input embedding plus GGUF overhead, and FP16 full-context KV. The audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/Meta-Llama-3.1-8B-Instruct-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Meta-Llama-3.1-8B-Instruct-GGUF" }, { "label": "Selected Q2_K GGUF artifact", "url": "https://huggingface.co/MaziyarPanahi/Meta-Llama-3.1-8B-Instruct-GGUF/resolve/1f301d86d760b435a11a56de3863bc0121bfb98f/Meta-Llama-3.1-8B-Instruct.Q2_K.gguf" }, { "label": "Gated base model profile", "url": "https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct" } ] }, { "id": "lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-8bit", "name": "Qwen3-Coder-30B-A3B-Instruct-MLX-8bit 8-bit", "short_name": "Qwen3 Coder 30B A3B Instruct MLX 8bit", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 163274, "tags": [ "transformers", "safetensors", "qwen3_moe", "text-generation", "mlx", "conversational", "base_model:Qwen/Qwen3-Coder-30B-A3B-Instruct", "base_model:quantized:Qwen/Qwen3-Coder-30B-A3B-Instruct", "license:apache-2.0", "endpoints_compatible", "8-bit", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_moe (Qwen3MoeForCausalLM)", "total_params_b": 30.532122624, "active_params_b": 3.042353139, "layers": 48, "attention_heads": 32, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "MLX 8-bit affine", "weight_bytes_per_param": 1.062506477112726, "resident_weight_gb": 32.440578048, "active_weight_gb": 3.232182272, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "notes": "Audited on 2026-07-06 from current Hugging Face API metadata, model card, pinned served MLX config, official BF16 base config comparison, safetensors index, and direct safetensors shard header range reads. Bound weights use exact stored MLX tensor payload bytes: 32.440578048 GB resident, 0.330612736 GB resident-only input embedding, 1.306996736 GB fixed decode traffic, and 0.240648192 GB per routed expert group." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-8bit" }, { "label": "Model config", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-8bit/raw/4285def38c15e8a2a01bf136a1a2349484003540/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-8bit" }, { "label": "Safetensors index", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-8bit/raw/4285def38c15e8a2a01bf136a1a2349484003540/model.safetensors.index.json" } ] }, { "id": "nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16", "name": "NVIDIA-Nemotron-Nano-12B-v2-VL-BF16 F32/BF16", "short_name": "NVIDIA Nemotron Nano 12B v2 VL BF16", "license": "other", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 174551, "tags": [ "transformers", "safetensors", "nvidia", "VLM", "image-text-to-text", "conversational", "arxiv:2501.14818", "arxiv:2511.03929", "license:other", "endpoints_compatible", "region:us" ], "architecture": { "type": "Hybrid vision-language model", "detail": "NemotronH_Nano_VL_V2; language: 62-layer Nemotron H with 28 Mamba, 6 attention, and 28 MLP-only layers; vision: C-RADIOv2-H plus projector", "total_params_b": 13.181860358, "active_params_b": 11.644155392, "layers": 62, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "hybrid", "weight_precision": "F32/BF16", "weight_bytes_per_param": 2.000000000910342, "resident_weight_gb": 26.363720728, "active_weight_gb": 23.288310784, "kv_alloc_gb_per_1k_tokens": 0.024576, "kv_read_gb_per_1k_tokens": 0.024576, "kv_alloc_fixed_gb": 0.149553152, "kv_read_fixed_gb": 0.149553152, "notes": "Manually corrected on 2026-07-06 from current Hugging Face API metadata, model card, served VL config, 12B parent text config comparison, custom runtime code, processor config, safetensors index, and direct range-read safetensors shard headers. Ordinary text decode charges exact language swept traffic and BF16 attention KV plus BF16-convolution/F32-SSM Mamba state; vision_model, mlp1 projector, and input embeddings are resident/prefill-side. The audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16" }, { "label": "Model config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16/raw/5d250e2e111dc5e1434131bdf3d590c27a878ade/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16" }, { "label": "Custom runtime code", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16/raw/5d250e2e111dc5e1434131bdf3d590c27a878ade/modeling.py" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16/resolve/5d250e2e111dc5e1434131bdf3d590c27a878ade/model.safetensors.index.json" } ] }, { "id": "LargitData/gemma-4-26b-a4b-it-fp8", "name": "gemma-4-26b-a4b-it-fp8 FP8", "short_name": "gemma 4 26b a4b it fp8", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 150515, "tags": [ "transformers", "safetensors", "gemma4", "image-text-to-text", "gemma-4", "vllm", "fp8", "fp8-dynamic", "compressed-tensors", "quantization", "h200", "nvidia-h200", "mixture-of-experts", "moe", "inference", "production-ready", "largitdata", "text-generation", "conversational", "en", "zh", "base_model:google/gemma-4-26B-A4B-it", "base_model:quantized:google/gemma-4-26B-A4B-it", "license:other", "endpoints_compatible", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "gemma4_text", "total_params_b": 25.8227, "active_params_b": 4, "layers": 30, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "FP8", "weight_bytes_per_param": 1, "resident_weight_gb": 27.161975452, "active_weight_gb": 4.57536262, "kv_alloc_gb_per_1k_tokens": 0.10496, "kv_read_gb_per_1k_tokens": 0.10496, "notes": "Manually corrected on 2026-07-06 from current Hugging Face API metadata, model card, served config, Google base config comparison, and direct safetensors shard header range reads. KV display fields show first-1K text-cache traffic for the documented vLLM --kv-cache-dtype fp8 launch; the audited profile supersedes these legacy adapter fields for production bounds and handles the 1024-token sliding-window cap." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/LargitData/gemma-4-26b-a4b-it-fp8" }, { "label": "Model config", "url": "https://huggingface.co/LargitData/gemma-4-26b-a4b-it-fp8/raw/2398f3b38d1b2e56d14498f7da3a74bada8cddff/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/LargitData/gemma-4-26b-a4b-it-fp8" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/LargitData/gemma-4-26b-a4b-it-fp8/raw/2398f3b38d1b2e56d14498f7da3a74bada8cddff/model.safetensors.index.json" } ] }, { "id": "MaziyarPanahi/Mixtral-8x22B-v0.1-GGUF", "name": "Mixtral-8x22B-v0.1 IQ1_M GGUF", "short_name": "Mixtral 8x22B v0.1", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 131580, "tags": [ "transformers", "gguf", "mixtral", "text-generation", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "16-bit", "GGUF", "moe", "fr", "en", "es", "it", "de", "base_model:v2ray/Mixtral-8x22B-v0.1", "base_model:quantized:v2ray/Mixtral-8x22B-v0.1", "license:apache-2.0", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "mixtral (MixtralForCausalLM)", "total_params_b": 140.6206, "active_params_b": 35.1551, "layers": 56, "attention_heads": 48, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 65536, "routed_experts": 8, "routed_experts_per_token": 2, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "GGUF IQ1_M", "weight_bytes_per_param": 0.23277094385685712, "resident_weight_gb": 32.732397728, "active_weight_gb": 10.00845312, "kv_alloc_gb_per_1k_tokens": 0.229376, "kv_read_gb_per_1k_tokens": 0.229376, "notes": "Manually corrected on 2026-07-06 from current Hugging Face API metadata, model card, target config, base config, linked-object HEAD checks, and direct selected-IQ1_M GGUF header range read. The audited profile supersedes these legacy adapter fields for production bounds and uses expected-distinct MoE expert traffic." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/Mixtral-8x22B-v0.1-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/MaziyarPanahi/Mixtral-8x22B-v0.1-GGUF/raw/053abe826eb9c5b1794e97bba22f0c50910e54e5/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/v2ray/Mixtral-8x22B-v0.1" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Mixtral-8x22B-v0.1-GGUF" }, { "label": "Selected IQ1_M GGUF header audit", "url": "https://huggingface.co/MaziyarPanahi/Mixtral-8x22B-v0.1-GGUF/resolve/053abe826eb9c5b1794e97bba22f0c50910e54e5/Mixtral-8x22B-v0.1.IQ1_M.gguf" } ] }, { "id": "lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-5bit", "name": "Qwen3-Coder-30B-A3B-Instruct-MLX-5bit 5-bit", "short_name": "Qwen3 Coder 30B A3B Instruct MLX 5bit", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 160568, "tags": [ "transformers", "safetensors", "qwen3_moe", "text-generation", "mlx", "conversational", "base_model:Qwen/Qwen3-Coder-30B-A3B-Instruct", "base_model:quantized:Qwen/Qwen3-Coder-30B-A3B-Instruct", "license:apache-2.0", "endpoints_compatible", "5-bit", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_moe (Qwen3MoeForCausalLM)", "total_params_b": 30.5321, "active_params_b": 3, "layers": 48, "attention_heads": 32, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "5-bit", "weight_bytes_per_param": 0.6875090679578164, "resident_weight_gb": 20.991111168, "active_weight_gb": 2.09156096, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "notes": "Manually corrected on 2026-07-06 from current Hugging Face API metadata, model card, served MLX config, base config comparison, and direct safetensors shard header range reads. The audited profile supersedes these legacy adapter fields for production bounds and uses exact stored MLX tensor bytes plus expected-distinct MoE expert traffic." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-5bit" }, { "label": "Model config", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-5bit/raw/a7d91f62488598d4725c5363a73d0a8f0e1b9329/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-5bit" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-5bit/raw/a7d91f62488598d4725c5363a73d0a8f0e1b9329/model.safetensors.index.json" } ] }, { "id": "stepfun-ai/step3", "name": "step3 BF16", "short_name": "step3", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 164103, "tags": [ "transformers", "safetensors", "step3_vl", "text-generation", "image-text-to-text", "conversational", "custom_code", "arxiv:2507.19427", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "Sparse MoE multimodal transformer", "detail": "step3_vl with Step3TextForCausalLM MoE language model", "total_params_b": 320.969912832, "active_params_b": 37.91798272, "layers": 61, "attention_heads": 64, "kv_heads": 1, "head_dim": 256, "max_context_tokens": 65536, "routed_experts": 48, "routed_experts_per_token": 3, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 641.939825664, "active_weight_gb": 75.83596544, "kv_alloc_gb_per_1k_tokens": 0.062464, "kv_read_gb_per_1k_tokens": 0.062464, "notes": "Manually audited on 2026-07-07 from live Hugging Face API metadata, model card, deployment guide, pinned config, remote modeling code, safetensors index, and direct shard header range reads. Step3 is a BF16 sparse MoE VLM with one KV head; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/stepfun-ai/step3" }, { "label": "Model config", "url": "https://huggingface.co/stepfun-ai/step3/raw/e2d67fc936dc60f8f606f2429728f1dea503e10d/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/stepfun-ai/step3" }, { "label": "Deployment guide", "url": "https://huggingface.co/stepfun-ai/step3/raw/e2d67fc936dc60f8f606f2429728f1dea503e10d/docs/deploy_guidance.md" }, { "label": "Remote modeling code", "url": "https://huggingface.co/stepfun-ai/step3/raw/e2d67fc936dc60f8f606f2429728f1dea503e10d/modeling_step3.py" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/stepfun-ai/step3/resolve/e2d67fc936dc60f8f606f2429728f1dea503e10d/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen1.5-0.5B-Chat", "name": "Qwen1.5-0.5B-Chat BF16", "short_name": "Qwen1.5 0.5B", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 179717, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "chat", "conversational", "en", "arxiv:2309.16609", "license:other", "text-generation-inference", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 0.619570176, "active_params_b": 0.463987712, "layers": 24, "attention_heads": 16, "kv_heads": 16, "head_dim": 64, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 1.239140352, "active_weight_gb": 0.927975424, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "notes": "Manually audited on 2026-07-07 from live Hugging Face API metadata, model card, pinned config, tokenizer config, and direct safetensors header range reads. Bound weights use exact BF16 stored bytes and separate the resident input embedding from ordinary cached text-decode traffic." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen1.5-0.5B-Chat" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen1.5-0.5B-Chat/raw/4d14e384a4b037942bb3f3016665157c8bcb70ea/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen1.5-0.5B-Chat" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/Qwen/Qwen1.5-0.5B-Chat/resolve/4d14e384a4b037942bb3f3016665157c8bcb70ea/model.safetensors" } ] }, { "id": "lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-6bit", "name": "Qwen3-Coder-30B-A3B-Instruct-MLX-6bit 6-bit", "short_name": "Qwen3 Coder 30B A3B Instruct MLX 6bit", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 159385, "tags": [ "transformers", "safetensors", "qwen3_moe", "text-generation", "mlx", "conversational", "base_model:Qwen/Qwen3-Coder-30B-A3B-Instruct", "base_model:quantized:Qwen/Qwen3-Coder-30B-A3B-Instruct", "license:apache-2.0", "endpoints_compatible", "6-bit", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_moe (Qwen3MoeForCausalLM)", "total_params_b": 30.5321, "active_params_b": 3, "layers": 48, "attention_heads": 32, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "6-bit", "weight_bytes_per_param": 0.8125082043427864, "resident_weight_gb": 24.807600128, "active_weight_gb": 2.471768064, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "notes": "Manually corrected on 2026-07-06 from current Hugging Face API metadata, model card, served MLX config, base config comparison, and direct safetensors shard header range reads. The audited profile supersedes these legacy adapter fields for production bounds and uses exact stored MLX tensor bytes plus expected-distinct MoE expert traffic." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-6bit" }, { "label": "Model config", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-6bit/raw/0cbf9904cf3b9a8a987e5d2cf59099464e7aa9e5/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-6bit" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-MLX-6bit/raw/0cbf9904cf3b9a8a987e5d2cf59099464e7aa9e5/model.safetensors.index.json" } ] }, { "id": "MaziyarPanahi/Qwen2.5-7B-Instruct-GGUF", "name": "Qwen2.5-7B-Instruct-GGUF GGUF FP16", "short_name": "Qwen2.5 7B Instruct", "license": "unknown", "hf_pipeline_tag": "text-generation", "hf_downloads": 124355, "tags": [ "gguf", "mistral", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:Qwen/Qwen2.5-7B-Instruct", "base_model:quantized:Qwen/Qwen2.5-7B-Instruct", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 GGUF (Qwen2ForCausalLM)", "total_params_b": 7.615616512, "active_params_b": 7.070619136, "layers": 28, "attention_heads": 28, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF FP16", "weight_bytes_per_param": 2, "resident_weight_gb": 15.237853632, "active_weight_gb": 14.141904896, "kv_alloc_gb_per_1k_tokens": 0.028672, "kv_read_gb_per_1k_tokens": 0.028672, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, model card, audited BF16 base profile comparison, linked GGUF file sizes, stale config check, and direct selected FP16 GGUF tensor-index range read. The authoritative bounds profile uses Qwen2.5-7B-Instruct.fp16.gguf because it matches API gguf.totalFileSize: 15.237853632 GB resident and 14.141904896 GB ordinary text-decode swept traffic after excluding token_embd.weight." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/Qwen2.5-7B-Instruct-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/MaziyarPanahi/Qwen2.5-7B-Instruct-GGUF/raw/113085215169f69e309fa57d351d7e2b06e1350e/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Qwen2.5-7B-Instruct-GGUF" }, { "label": "Selected GGUF tensor-index audit", "url": "https://huggingface.co/MaziyarPanahi/Qwen2.5-7B-Instruct-GGUF/resolve/113085215169f69e309fa57d351d7e2b06e1350e/Qwen2.5-7B-Instruct.fp16.gguf" } ] }, { "id": "MaziyarPanahi/Phi-4-mini-instruct-GGUF", "name": "Phi-4-mini-instruct-GGUF GGUF FP16", "short_name": "Phi 4 mini instruct", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 126346, "tags": [ "gguf", "mistral", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:microsoft/Phi-4-mini-instruct", "base_model:quantized:microsoft/Phi-4-mini-instruct", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "phi3 GGUF (Phi3ForCausalLM)", "total_params_b": 3.836021856, "active_params_b": 3.836021856, "layers": 32, "attention_heads": 24, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF FP16", "weight_bytes_per_param": 2, "resident_weight_gb": 7.680694592, "active_weight_gb": 7.672443264, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, model card, audited BF16 base profile comparison, linked GGUF file sizes, stale config check, and direct selected FP16 GGUF tensor-index range read. The authoritative bounds profile uses Phi-4-mini-instruct.fp16.gguf because it matches API gguf.totalFileSize: 7.680694592 GB resident and 7.672443264 GB ordinary text-decode swept traffic." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/Phi-4-mini-instruct-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/MaziyarPanahi/Phi-4-mini-instruct-GGUF/raw/177ca4b8b568a3f865c34e21e914aeb553da19cc/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/microsoft/Phi-4-mini-instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Phi-4-mini-instruct-GGUF" }, { "label": "Selected GGUF tensor-index audit", "url": "https://huggingface.co/MaziyarPanahi/Phi-4-mini-instruct-GGUF/resolve/177ca4b8b568a3f865c34e21e914aeb553da19cc/Phi-4-mini-instruct.fp16.gguf" } ] }, { "id": "MaziyarPanahi/Llama-3.3-70B-Instruct-GGUF", "name": "Llama-3.3-70B-Instruct-GGUF GGUF Q2_K", "short_name": "Llama 3.3 70B Instruct", "license": "llama3.3", "hf_pipeline_tag": "text-generation", "hf_downloads": 125907, "tags": [ "gguf", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:meta-llama/Llama-3.3-70B-Instruct", "base_model:quantized:meta-llama/Llama-3.3-70B-Instruct", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "llama GGUF (LlamaForCausalLM)", "total_params_b": 70.55370656, "active_params_b": 69.503033408, "layers": 80, "attention_heads": 64, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q2_K", "weight_bytes_per_param": 0.373830302134023, "resident_weight_gb": 26.37511344, "active_weight_gb": 26.022494464, "kv_alloc_gb_per_1k_tokens": 0.32768, "kv_read_gb_per_1k_tokens": 0.32768, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, model card, gated base profile, audited Casper Hansen AWQ config comparison, linked GGUF file sizes, missing config check, and direct selected Q2_K GGUF tensor-index range read. The authoritative bounds profile uses Llama-3.3-70B-Instruct.Q2_K.gguf because it matches API gguf.totalFileSize: 26.37511344 GB resident and 26.022494464 GB ordinary text-decode swept traffic after excluding token_embd.weight." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/Llama-3.3-70B-Instruct-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct/raw/6f6073b423013f6a7d4d9f39144961bfbfbc386b/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Llama-3.3-70B-Instruct-GGUF" }, { "label": "Selected GGUF tensor-index audit", "url": "https://huggingface.co/MaziyarPanahi/Llama-3.3-70B-Instruct-GGUF/resolve/b09bf396ffd1df6c4c46d66c932f2c3cc2b7b21c/Llama-3.3-70B-Instruct.Q2_K.gguf" } ] }, { "id": "cyankiwi/Qwen3-VL-30B-A3B-Instruct-AWQ-4bit", "name": "Qwen3-VL-30B-A3B-Instruct-AWQ-4bit 4-bit", "short_name": "Qwen3 VL 30B A3B Instruct AWQ 4bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 149879, "tags": [ "transformers", "safetensors", "qwen3_vl_moe", "image-text-to-text", "conversational", "arxiv:2505.09388", "arxiv:2502.13923", "arxiv:2409.12191", "arxiv:2308.12966", "base_model:Qwen/Qwen3-VL-30B-A3B-Instruct", "base_model:quantized:Qwen/Qwen3-VL-30B-A3B-Instruct", "license:apache-2.0", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_vl_moe_text", "total_params_b": 31.070754032, "active_params_b": 3, "layers": 48, "attention_heads": 32, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "compressed-tensors AWQ 4-bit + BF16/I64 side tensors", "weight_bytes_per_param": 0.61681380800292, "resident_weight_gb": 19.164870112, "active_weight_gb": 2.17676288, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "notes": "Audited from Hugging Face metadata on 2026-07-07; downloads=149879; data quality: config, base config, API params, recipe, safetensors headers. Bound weights use exact stored compressed-tensors I32/BF16/I64 tensor byte groups." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/cyankiwi/Qwen3-VL-30B-A3B-Instruct-AWQ-4bit" }, { "label": "Model config", "url": "https://huggingface.co/cyankiwi/Qwen3-VL-30B-A3B-Instruct-AWQ-4bit/raw/3ab8f1af7e50ee8e54eb12b7bf9eeedc5a84184e/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3-VL-30B-A3B-Instruct-AWQ-4bit" }, { "label": "Safetensors index", "url": "https://huggingface.co/cyankiwi/Qwen3-VL-30B-A3B-Instruct-AWQ-4bit/resolve/3ab8f1af7e50ee8e54eb12b7bf9eeedc5a84184e/model.safetensors.index.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct/raw/9c4b90e1e4ba969fd3b5378b57d966d725f1b86c/config.json" } ] }, { "id": "MaziyarPanahi/Yi-Coder-1.5B-Chat-GGUF", "name": "Yi-Coder-1.5B-Chat-GGUF GGUF Q2_K", "short_name": "Yi Coder 1.5B Chat", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 124754, "tags": [ "gguf", "mistral", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:01-ai/Yi-Coder-1.5B-Chat", "base_model:quantized:01-ai/Yi-Coder-1.5B-Chat", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "llama GGUF (LlamaForCausalLM)", "total_params_b": 1.47649536, "active_params_b": 1.34542336, "layers": 24, "attention_heads": 16, "kv_heads": 16, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q2_K", "weight_bytes_per_param": 0.42986926013773585, "resident_weight_gb": 0.634699968, "active_weight_gb": 0.590200832, "kv_alloc_gb_per_1k_tokens": 0.196608, "kv_read_gb_per_1k_tokens": 0.196608, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, model card, pinned base config comparison, linked GGUF file sizes, stale config check, and direct selected Q2_K GGUF tensor-index range read. The authoritative bounds profile uses Yi-Coder-1.5B-Chat.Q2_K.gguf for the catalog's 2-bit row: 0.634699968 GB resident and 0.590200832 GB ordinary text-decode swept traffic after excluding token_embd.weight. The live API gguf.totalFileSize currently points to the FP16 sibling, so selected-file header bytes are authoritative." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/Yi-Coder-1.5B-Chat-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/MaziyarPanahi/Yi-Coder-1.5B-Chat-GGUF/raw/cc8525f15600e412829215092967f636b654913a/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Yi-Coder-1.5B-Chat-GGUF" }, { "label": "Selected GGUF tensor-index audit", "url": "https://huggingface.co/MaziyarPanahi/Yi-Coder-1.5B-Chat-GGUF/resolve/cc8525f15600e412829215092967f636b654913a/Yi-Coder-1.5B-Chat.Q2_K.gguf" }, { "label": "Base model card", "url": "https://huggingface.co/01-ai/Yi-Coder-1.5B-Chat" }, { "label": "Base model config", "url": "https://huggingface.co/01-ai/Yi-Coder-1.5B-Chat/raw/92fdd1b2f1539ac990e7f4a921db5601da2f0299/config.json" } ] }, { "id": "MaziyarPanahi/DeepSeek-R1-0528-Qwen3-8B-GGUF", "name": "DeepSeek-R1-0528-Qwen3-8B-GGUF GGUF Q2_K", "short_name": "DeepSeek R1 0528 Qwen3 8B", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 124510, "tags": [ "gguf", "mistral", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:deepseek-ai/DeepSeek-R1-0528-Qwen3-8B", "base_model:quantized:deepseek-ai/DeepSeek-R1-0528-Qwen3-8B", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 GGUF (Qwen3ForCausalLM)", "total_params_b": 8.19073536, "active_params_b": 7.568405504, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q2_K", "weight_bytes_per_param": 0.40066386810963944, "resident_weight_gb": 3.281731712, "active_weight_gb": 3.071574016, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, model card, audited BF16 base profile, linked GGUF file sizes, stale config check, and direct selected Q2_K GGUF tensor-index range read. The authoritative bounds profile uses DeepSeek-R1-0528-Qwen3-8B.Q2_K.gguf for the catalog's 2-bit row: 3.281731712 GB resident and 3.071574016 GB ordinary text-decode swept traffic after excluding token_embd.weight. The live API gguf.totalFileSize currently points to the FP16 sibling, so selected-file header bytes are authoritative." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/DeepSeek-R1-0528-Qwen3-8B-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/MaziyarPanahi/DeepSeek-R1-0528-Qwen3-8B-GGUF/raw/3d7c2a7c46f7cd566b7318e49986dda9a6d7d190/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/DeepSeek-R1-0528-Qwen3-8B-GGUF" }, { "label": "Selected GGUF tensor-index audit", "url": "https://huggingface.co/MaziyarPanahi/DeepSeek-R1-0528-Qwen3-8B-GGUF/resolve/3d7c2a7c46f7cd566b7318e49986dda9a6d7d190/DeepSeek-R1-0528-Qwen3-8B.Q2_K.gguf" }, { "label": "Base model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-0528-Qwen3-8B" }, { "label": "Base model config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-R1-0528-Qwen3-8B/raw/6e8885a6ff5c1dc5201574c8fd700323f23c25fa/config.json" } ] }, { "id": "RedHatAI/Qwen3.5-122B-A10B-NVFP4", "name": "Qwen3.5-122B-A10B-NVFP4 NVFP4", "short_name": "Qwen3.5 122B A10B NVFP4", "license": "unknown", "hf_pipeline_tag": "unknown", "hf_downloads": 246347, "tags": [ "safetensors", "qwen3_5_moe", "qwen", "nvfp4", "vllm", "compressed-tensors", "base_model:Qwen/Qwen3.5-122B-A10B", "base_model:quantized:Qwen/Qwen3.5-122B-A10B", "8-bit", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_5_moe_text", "total_params_b": 125.086497008, "active_params_b": 10, "layers": 48, "attention_heads": 32, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "compressed-tensors NVFP4/FP8/BF16/F32", "weight_bytes_per_param": 0.5, "resident_weight_gb": 76.419752928, "active_weight_gb": 10.79982336, "kv_alloc_gb_per_1k_tokens": 0.024576, "kv_read_gb_per_1k_tokens": 0.024576, "kv_alloc_fixed_gb": 0.154533888, "kv_read_fixed_gb": 0.154533888, "expert_param_b": 0.50961024, "fixed_param_b": 17.5227648, "shared_expert_param_b": 0, "notes": "Manually corrected on 2026-07-07 from current Hugging Face API metadata, pinned compressed-tensors NVFP4 config, recipe, BF16 base config comparison, safetensors index, and direct range-read safetensors shard headers. Legacy fallback values mirror the audited profile: exact stored resident bytes, fixed always-on hybrid/shared-expert traffic, expected distinct routed experts, BF16 attention KV, and fixed BF16-convolution/F32-SSM DeltaNet recurrent state. The target repo does not currently declare license or pipeline metadata. The authoritative profile supersedes the generated 62.5433 GB resident estimate, the missing shared-expert count, and the all-layers full-context KV estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/RedHatAI/Qwen3.5-122B-A10B-NVFP4" }, { "label": "Model config", "url": "https://huggingface.co/RedHatAI/Qwen3.5-122B-A10B-NVFP4/raw/49d19c108259a21450c40b8af38828b0a97390d8/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/RedHatAI/Qwen3.5-122B-A10B-NVFP4" }, { "label": "Compression recipe", "url": "https://huggingface.co/RedHatAI/Qwen3.5-122B-A10B-NVFP4/raw/49d19c108259a21450c40b8af38828b0a97390d8/recipe.yaml" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B" }, { "label": "Base config", "url": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B/raw/dc4d348443bc740c68e2d77492492c11606384d5/config.json" }, { "label": "Safetensors index", "url": "https://huggingface.co/RedHatAI/Qwen3.5-122B-A10B-NVFP4/resolve/49d19c108259a21450c40b8af38828b0a97390d8/model.safetensors.index.json" } ] }, { "id": "deepreinforce-ai/Ornith-1.0-35B-GGUF", "name": "Ornith-1.0-35B-GGUF GGUF BF16", "short_name": "Ornith 1.0 35B", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 436780, "tags": [ "transformers", "gguf", "text-generation", "license:mit", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "MoE transformer", "detail": "qwen35moe / qwen3_5_moe_text", "total_params_b": 34.660610688, "active_params_b": 3, "layers": 40, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "GGUF BF16/F32", "weight_bytes_per_param": 2, "resident_weight_gb": 69.3766368, "active_weight_gb": 5.937285632, "kv_alloc_gb_per_1k_tokens": 0.02048, "kv_read_gb_per_1k_tokens": 0.02048, "kv_alloc_fixed_gb": 0.06488064, "kv_read_fixed_gb": 0.06488064, "expert_param_b": 0.12582912, "fixed_param_b": 1.962009856, "shared_expert_param_b": 0, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, model card, non-GGUF Ornith config/API metadata, linked GGUF file sizes, and direct selected BF16 GGUF tensor-index range read. The authoritative bounds profile uses ornith-1.0-35b-bf16.gguf because it matches API gguf.totalFileSize: 69.376636800 GB resident and 5.937285632 GB ordinary single-session swept traffic before KV/state reads. The audited profile supersedes the generated dense 35B, 2k-context, zero-KV placeholder." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/deepreinforce-ai/Ornith-1.0-35B-GGUF" }, { "label": "Non-GGUF Ornith config", "url": "https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B/raw/5df2ed3f675c7beaa490328cc70bb573b65fb660/config.json" }, { "label": "Non-GGUF Ornith API metadata", "url": "https://huggingface.co/api/models/deepreinforce-ai/Ornith-1.0-35B" }, { "label": "Selected GGUF header audit", "url": "https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B-GGUF/resolve/c2e1703039380de4ce6820e97afd185682d3c16c/ornith-1.0-35b-bf16.gguf" } ] }, { "id": "MaziyarPanahi/Mistral-Nemo-Instruct-2407-GGUF", "name": "Mistral-Nemo-Instruct-2407-GGUF GGUF F16", "short_name": "Mistral Nemo Instruct 2407", "license": "unknown", "hf_pipeline_tag": "text-generation", "hf_downloads": 123676, "tags": [ "transformers", "gguf", "mistral", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:mistralai/Mistral-Nemo-Instruct-2407", "base_model:quantized:mistralai/Mistral-Nemo-Instruct-2407", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "llama GGUF / MistralForCausalLM", "total_params_b": 12.2477824, "active_params_b": 11.57669376, "layers": 40, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 1024000, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF F16/F32", "weight_bytes_per_param": 2, "resident_weight_gb": 24.504276704, "active_weight_gb": 23.15421696, "kv_alloc_gb_per_1k_tokens": 0.16384, "kv_read_gb_per_1k_tokens": 0.16384, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, model card, base Mistral Nemo config/API metadata, linked GGUF file sizes, and direct selected F16 GGUF tensor-index range read. The authoritative bounds profile uses Mistral-Nemo-Instruct-2407.fp16.gguf because it matches API gguf.totalFileSize: 24.504276704 GB resident and 23.154216960 GB ordinary text-decode swept traffic after excluding token_embd.weight plus GGUF metadata/header/file overhead. The audited profile supersedes the generated 2-bit, 2k-context, zero-KV placeholder." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/Mistral-Nemo-Instruct-2407-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/MaziyarPanahi/Mistral-Nemo-Instruct-2407-GGUF/raw/eba4e7492de28b8ab2ff44b0bb819004181b3db4/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Mistral-Nemo-Instruct-2407-GGUF" }, { "label": "Base model card", "url": "https://huggingface.co/mistralai/Mistral-Nemo-Instruct-2407" }, { "label": "Base config", "url": "https://huggingface.co/mistralai/Mistral-Nemo-Instruct-2407/raw/04d8a90549d23fc6bd7f642064003592df51e9b3/config.json" }, { "label": "Selected GGUF header audit", "url": "https://huggingface.co/MaziyarPanahi/Mistral-Nemo-Instruct-2407-GGUF/resolve/eba4e7492de28b8ab2ff44b0bb819004181b3db4/Mistral-Nemo-Instruct-2407.fp16.gguf" } ] }, { "id": "allenai/Molmo2-O-7B", "name": "Molmo2-O-7B F32", "short_name": "Molmo2 O 7B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 161905, "tags": [ "transformers", "safetensors", "molmo2", "image-text-to-text", "multimodal", "olmo", "molmo", "conversational", "custom_code", "en", "dataset:allenai/Molmo2-Cap", "dataset:allenai/Molmo2-VideoCapQA", "dataset:allenai/Molmo2-VideoSubtitleQA", "dataset:allenai/Molmo2-AskModelAnything", "dataset:allenai/Molmo2-VideoPoint", "dataset:allenai/Molmo2-VideoTrack", "dataset:allenai/Molmo2-MultiImageQA", "dataset:allenai/Molmo2-SynMultiImageQA", "dataset:allenai/Molmo2-MultiImagePoint", "base_model:allenai/Olmo-3-7B-Instruct", "base_model:finetune:allenai/Olmo-3-7B-Instruct", "license:apache-2.0", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "molmo2_text (Molmo2ForConditionalGeneration, Olmo-3-like)", "total_params_b": 7.760786896, "active_params_b": 6.887272448, "layers": 32, "attention_heads": 32, "kv_heads": 32, "head_dim": 128, "max_context_tokens": 65536, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "F32", "weight_bytes_per_param": 4, "resident_weight_gb": 31.043147584, "active_weight_gb": 27.549089792, "kv_alloc_gb_per_1k_tokens": 1.048576, "kv_read_gb_per_1k_tokens": 1.048576, "notes": "Audited 2026-07-07 from pinned config and range-read safetensors shard headers; downloads=161905. Resident storage is the full F32 multimodal package. Active ordinary text-decode traffic excludes input embeddings and vision backbone, and KV is full-context F32 for 32 text layers x 32 KV heads." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/allenai/Molmo2-O-7B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/allenai/Molmo2-O-7B" }, { "label": "Model config", "url": "https://huggingface.co/allenai/Molmo2-O-7B/raw/784410650d12be9bc086118fdefa32d2c3bced86/config.json" }, { "label": "Custom modeling code", "url": "https://huggingface.co/allenai/Molmo2-O-7B/raw/784410650d12be9bc086118fdefa32d2c3bced86/modeling_molmo2.py" }, { "label": "Safetensors index", "url": "https://huggingface.co/allenai/Molmo2-O-7B/raw/784410650d12be9bc086118fdefa32d2c3bced86/model.safetensors.index.json" } ] }, { "id": "MaziyarPanahi/Llama-3-8B-Instruct-32k-v0.1-GGUF", "name": "Llama-3-8B-Instruct-32k-v0.1-GGUF FP16", "short_name": "Llama 3 8B Instruct 32k v0.1", "license": "unknown", "hf_pipeline_tag": "text-generation", "hf_downloads": 122730, "tags": [ "transformers", "gguf", "mistral", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "llama (GGUF header; package config says mistral)", "total_params_b": 8.030261248, "active_params_b": 7.504924672, "layers": 32, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 8192, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF FP16", "weight_bytes_per_param": 2, "resident_weight_gb": 16.068890944, "active_weight_gb": 15.010381824, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "notes": "Audited 2026-07-07 from live HF API metadata and a range-read of the selected FP16 GGUF header; downloads=122730. API gguf.totalFileSize selects Llama-3-8B-Instruct-32k-v0.1.fp16.gguf. The selected header records 8192 context despite the 32k repo name." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/Llama-3-8B-Instruct-32k-v0.1-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Llama-3-8B-Instruct-32k-v0.1-GGUF" }, { "label": "Package config conflict check", "url": "https://huggingface.co/MaziyarPanahi/Llama-3-8B-Instruct-32k-v0.1-GGUF/raw/4251e4b7ec372ba0de6046137272a709e349f442/config.json" }, { "label": "Selected FP16 GGUF artifact", "url": "https://huggingface.co/MaziyarPanahi/Llama-3-8B-Instruct-32k-v0.1-GGUF/resolve/4251e4b7ec372ba0de6046137272a709e349f442/Llama-3-8B-Instruct-32k-v0.1.fp16.gguf" } ] }, { "id": "MaziyarPanahi/gemma-3-1b-it-GGUF", "name": "gemma-3-1b-it-GGUF FP16", "short_name": "gemma 3 1b it", "license": "gemma", "hf_pipeline_tag": "text-generation", "hf_downloads": 122255, "tags": [ "gguf", "mistral", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:google/gemma-3-1b-it", "base_model:quantized:google/gemma-3-1b-it", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "gemma3 (GGUF header; package config says mistral)", "total_params_b": 0.999885952, "active_params_b": 0.999885952, "layers": 26, "attention_heads": 4, "kv_heads": 1, "head_dim": 256, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF FP16", "weight_bytes_per_param": 2, "resident_weight_gb": 2.0065736, "active_weight_gb": 2.000040448, "kv_alloc_gb_per_1k_tokens": 0.015630336, "kv_read_gb_per_1k_tokens": 0.015630336, "notes": "Audited 2026-07-07 from live HF API metadata and a range-read of the selected FP16 GGUF header; downloads=122255. API gguf.totalFileSize selects gemma-3-1b-it.fp16.gguf. KV display fields show first-1K text-cache traffic; the audited profile supersedes these legacy adapter fields for production bounds and handles the 512-token local sliding-window cap." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/gemma-3-1b-it-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/gemma-3-1b-it-GGUF" }, { "label": "Package config conflict check", "url": "https://huggingface.co/MaziyarPanahi/gemma-3-1b-it-GGUF/raw/9aea24a60427bba55a903a1c4b526842d40fad3e/config.json" }, { "label": "Base model API metadata", "url": "https://huggingface.co/api/models/google/gemma-3-1b-it" }, { "label": "Selected FP16 GGUF artifact", "url": "https://huggingface.co/MaziyarPanahi/gemma-3-1b-it-GGUF/resolve/9aea24a60427bba55a903a1c4b526842d40fad3e/gemma-3-1b-it.fp16.gguf" } ] }, { "id": "QuantTrio/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2-AWQ", "name": "Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2-AWQ 4-bit", "short_name": "Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 216646, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "vLLM", "AWQ", "conversational", "base_model:Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2", "base_model:quantized:Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2", "license:apache-2.0", "endpoints_compatible", "4-bit", "awq", "region:us" ], "architecture": { "type": "Dense multimodal transformer", "detail": "qwen3_5_text (AWQ)", "total_params_b": 27.781427952, "active_params_b": 25.624600064, "layers": 64, "attention_heads": 24, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "AWQ 4-bit", "weight_bytes_per_param": 0.7865993885467936, "resident_weight_gb": 21.85285424, "active_weight_gb": 17.539198464, "kv_alloc_gb_per_1k_tokens": 0.065536, "kv_read_gb_per_1k_tokens": 0.065536, "kv_alloc_fixed_gb": 0.154927104, "kv_read_fixed_gb": 0.154927104, "notes": "Manual correction on 2026-07-07 from current Hugging Face API metadata, pinned model card, served AWQ config, Jackrong base config comparison, safetensors index, direct safetensors header range reads, and Qwen3.5 runtime review. Resident weights are exact stored tensor payloads; active traffic is ordinary text decode through language layers/norm and lm_head; resident-only input embedding, visual, and MTP tensors are excluded from per-token swept traffic. The audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/QuantTrio/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2-AWQ" }, { "label": "Pinned AWQ model config", "url": "https://huggingface.co/QuantTrio/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2-AWQ/raw/c633586811ef0e1d5938809364d68f8a5e401fe0/config.json" }, { "label": "Jackrong base model config", "url": "https://huggingface.co/Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2/raw/main/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/QuantTrio/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2-AWQ" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/QuantTrio/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-v2-AWQ/resolve/c633586811ef0e1d5938809364d68f8a5e401fe0/model.safetensors.index.json" }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py" } ] }, { "id": "MaziyarPanahi/Meta-Llama-3.1-70B-Instruct-GGUF", "name": "Meta-Llama-3.1-70B-Instruct-GGUF GGUF IQ1_M", "short_name": "Meta Llama 3.1 70B Instruct", "license": "llama3.1", "hf_pipeline_tag": "text-generation", "hf_downloads": 122003, "tags": [ "gguf", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "en", "de", "fr", "it", "pt", "hi", "es", "th", "base_model:meta-llama/Llama-3.1-70B-Instruct", "base_model:quantized:meta-llama/Llama-3.1-70B-Instruct", "license:llama3.1", "region:us", "imatrix", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "llama GGUF (LlamaForCausalLM)", "total_params_b": 70.553706496, "active_params_b": 69.503033344, "layers": 80, "attention_heads": 64, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF IQ1_M", "weight_bytes_per_param": 0.23742475229064963, "resident_weight_gb": 16.751196288, "active_weight_gb": 16.39858176, "kv_alloc_gb_per_1k_tokens": 0.32768, "kv_read_gb_per_1k_tokens": 0.32768, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, model card, gated base profile, audited Hugging Quants AWQ config comparison, linked GGUF file sizes, missing config check, and direct selected IQ1_M GGUF tensor-index range read. The authoritative bounds profile uses Meta-Llama-3.1-70B-Instruct.IQ1_M.gguf because it matches API gguf.totalFileSize: 16.751196288 GB resident and 16.398581760 GB ordinary text-decode swept traffic after excluding token_embd.weight." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/Meta-Llama-3.1-70B-Instruct-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Meta-Llama-3.1-70B-Instruct-GGUF" }, { "label": "Gated base model profile", "url": "https://huggingface.co/meta-llama/Llama-3.1-70B-Instruct" }, { "label": "Hugging Quants AWQ config comparison", "url": "https://huggingface.co/hugging-quants/Meta-Llama-3.1-70B-Instruct-AWQ-INT4/raw/2123003760781134cfc31124aa6560a45b491fdf/config.json" }, { "label": "Selected IQ1_M GGUF tensor-index audit", "url": "https://huggingface.co/MaziyarPanahi/Meta-Llama-3.1-70B-Instruct-GGUF/resolve/c1c0c3d88f2fca6f18de9063917f9bc49a33365a/Meta-Llama-3.1-70B-Instruct.IQ1_M.gguf" } ] }, { "id": "unsloth/Qwen-AgentWorld-35B-A3B-GGUF", "name": "Qwen-AgentWorld-35B-A3B-GGUF GGUF MXFP4_MOE", "short_name": "Qwen AgentWorld 35B A3B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 349969, "tags": [ "transformers", "gguf", "qwen", "unsloth", "world-model", "agent", "environment-simulation", "text-generation", "dataset:Qwen/AgentWorldBench", "arxiv:2606.24597", "base_model:Qwen/Qwen-AgentWorld-35B-A3B", "base_model:quantized:Qwen/Qwen-AgentWorld-35B-A3B", "license:apache-2.0", "endpoints_compatible", "region:us", "imatrix", "conversational" ], "architecture": { "type": "MoE transformer", "detail": "qwen35moe / qwen3_5_moe_text", "total_params_b": 34.660610688, "active_params_b": 3, "layers": 40, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "GGUF MXFP4_MOE", "weight_bytes_per_param": 0.6252196083637568, "resident_weight_gb": 21.67049344, "active_weight_gb": 2.731002368, "kv_alloc_gb_per_1k_tokens": 0.02048, "kv_read_gb_per_1k_tokens": 0.02048, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, model card, Qwen AgentWorld base config, expanded linked-file size metadata, and a direct selected MXFP4_MOE GGUF tensor-index range read. The selected API artifact is Qwen-AgentWorld-35B-A3B-MXFP4_MOE.gguf at 21.670493440 GB resident; ordinary single-session swept traffic is 2.731002368 GB before KV/state reads. The audited profile adds the fixed DeltaNet state charge that these legacy per-1K KV fields cannot fully express." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/Qwen-AgentWorld-35B-A3B-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen-AgentWorld-35B-A3B-GGUF" }, { "label": "Pinned base model config", "url": "https://huggingface.co/Qwen/Qwen-AgentWorld-35B-A3B/raw/60d2b0434a53d2e62a7c00a489586815d94ebffb/config.json" }, { "label": "Selected GGUF header audit", "url": "https://huggingface.co/unsloth/Qwen-AgentWorld-35B-A3B-GGUF/resolve/3a305abf5cfd119ee999dfe929c433746edd8d63/Qwen-AgentWorld-35B-A3B-MXFP4_MOE.gguf" } ] }, { "id": "MaziyarPanahi/Llama-3.2-1B-Instruct-GGUF", "name": "Llama-3.2-1B-Instruct-GGUF GGUF F16", "short_name": "Llama 3.2 1B Instruct", "license": "llama3.2", "hf_pipeline_tag": "text-generation", "hf_downloads": 119953, "tags": [ "gguf", "facebook", "meta", "llama", "llama-3", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:meta-llama/Llama-3.2-1B-Instruct", "base_model:quantized:meta-llama/Llama-3.2-1B-Instruct", "license:llama3.2", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "llama", "total_params_b": 1.235814432, "active_params_b": 1.235814432, "layers": 16, "attention_heads": 32, "kv_heads": 8, "head_dim": 64, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF F16", "weight_bytes_per_param": 2.0064463788362685, "resident_weight_gb": 2.479595392, "active_weight_gb": 2.471764096, "kv_alloc_gb_per_1k_tokens": 0.032768, "kv_read_gb_per_1k_tokens": 0.032768, "notes": "Manual correction on 2026-07-07 from current Hugging Face API metadata, model card, repo config-stub check, base API metadata, gated base-config access check, expanded linked-file size metadata, and a direct range-read GGUF tensor index. The authoritative bounds profile targets Llama-3.2-1B-Instruct.fp16.gguf because API gguf.totalFileSize exactly matches that linked object: 2.479595392 GB resident, 2.471764096 GB swept tensor spans, and 0.007831296 GB resident-only GGUF metadata/header/file overhead. The repo-local config.json says mistral but is a stale 31-byte stub; the selected GGUF header records llama." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/Llama-3.2-1B-Instruct-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/MaziyarPanahi/Llama-3.2-1B-Instruct-GGUF/raw/b64ae94264258a3d7516a34a8c54928d32b19869/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Llama-3.2-1B-Instruct-GGUF" }, { "label": "Selected F16 GGUF artifact", "url": "https://huggingface.co/MaziyarPanahi/Llama-3.2-1B-Instruct-GGUF/resolve/b64ae94264258a3d7516a34a8c54928d32b19869/Llama-3.2-1B-Instruct.fp16.gguf" } ] }, { "id": "lmstudio-community/Qwen2.5-Coder-14B-Instruct-MLX-4bit", "name": "Qwen2.5-Coder-14B-Instruct-MLX-4bit MLX 4-bit affine", "short_name": "Qwen2.5 Coder 14B Instruct MLX 4bit", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 150668, "tags": [ "mlx", "safetensors", "qwen2", "code", "codeqwen", "chat", "qwen", "qwen-coder", "text-generation", "conversational", "en", "base_model:Qwen/Qwen2.5-Coder-14B-Instruct", "base_model:quantized:Qwen/Qwen2.5-Coder-14B-Instruct", "license:apache-2.0", "4-bit", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 14.770033664, "active_params_b": 13.991465984, "layers": 48, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "MLX 4-bit affine", "weight_bytes_per_param": 0.5625818218852775, "resident_weight_gb": 8.309352448, "active_weight_gb": 7.871408128, "kv_alloc_gb_per_1k_tokens": 0.196608, "kv_read_gb_per_1k_tokens": 0.196608, "notes": "Manual correction on 2026-07-07 from current Hugging Face API metadata, model card, served MLX config, audited Qwen2.5-Coder 14B Instruct BF16 profile/config comparison, safetensors index, and direct safetensors header range reads. Bound weights separate the 8.309352448 GB resident MLX tensor package from the 7.871408128 GB ordinary text-decode swept subset; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/Qwen2.5-Coder-14B-Instruct-MLX-4bit" }, { "label": "Model config", "url": "https://huggingface.co/lmstudio-community/Qwen2.5-Coder-14B-Instruct-MLX-4bit/raw/4275f1d1fd379c8a5e8cc655c5a57ef03b912a29/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-14B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen2.5-Coder-14B-Instruct-MLX-4bit" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/lmstudio-community/Qwen2.5-Coder-14B-Instruct-MLX-4bit/raw/4275f1d1fd379c8a5e8cc655c5a57ef03b912a29/model.safetensors.index.json" } ] }, { "id": "batiai/Qwen3.6-27B-GGUF", "name": "Qwen3.6-27B-GGUF GGUF", "short_name": "Qwen3.6 27B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 154057, "tags": [ "llama.cpp", "gguf", "qwen", "qwen3.6", "dense", "quantized", "imatrix", "apple-silicon", "ollama", "batiai", "on-device", "agentic", "coding", "multimodal", "text-generation", "en", "ko", "ja", "zh", "base_model:Qwen/Qwen3.6-27B", "base_model:quantized:Qwen/Qwen3.6-27B", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense multimodal transformer", "detail": "qwen3_5_text (API-selected IQ3_XXS GGUF)", "total_params_b": 26.895998464, "active_params_b": 25.624600064, "layers": 64, "attention_heads": 24, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF IQ3_XXS", "weight_bytes_per_param": 0.4159120701532175, "resident_weight_gb": 11.1863704, "active_weight_gb": 10.629072896, "kv_alloc_gb_per_1k_tokens": 0.065536, "kv_read_gb_per_1k_tokens": 0.065536, "kv_alloc_fixed_gb": 0.154927104, "kv_read_fixed_gb": 0.154927104, "notes": "Manually audited on 2026-07-07 from HF API commit 8057b95492cfe844051386e1e4f04ee9c99f7b95, model card, pinned Qwen base config, linked GGUF file sizes, and a GGUF tensor-index range read. The authoritative bounds profile uses the API-selected Qwen-Qwen3.6-27B-IQ3_XXS.gguf artifact: 11.186370400 GB resident, 10.629072896 GB swept ordinary text traffic, 0.557297504 GB resident-only token embedding plus GGUF metadata/header/file overhead, and Qwen3.6 hybrid KV plus fixed DeltaNet state. Separate mmproj sidecars are not included unless explicitly loaded." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/batiai/Qwen3.6-27B-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.6-27B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/batiai/Qwen3.6-27B-GGUF" }, { "label": "Selected IQ3_XXS GGUF artifact", "url": "https://huggingface.co/batiai/Qwen3.6-27B-GGUF/resolve/8057b95492cfe844051386e1e4f04ee9c99f7b95/Qwen-Qwen3.6-27B-IQ3_XXS.gguf" } ] }, { "id": "MaziyarPanahi/gemma-2-2b-it-GGUF", "name": "gemma-2-2b-it-GGUF GGUF FP16", "short_name": "gemma 2 2b it", "license": "gemma", "hf_pipeline_tag": "text-generation", "hf_downloads": 119685, "tags": [ "transformers", "gguf", "gemma", "gemma2", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:google/gemma-2-2b-it", "base_model:quantized:google/gemma-2-2b-it", "license:gemma", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "gemma2", "total_params_b": 2.614341888, "active_params_b": 2.614341888, "layers": 26, "attention_heads": 8, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 8192, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF FP16", "weight_bytes_per_param": 2.0024978278586953, "resident_weight_gb": 5.235213952, "active_weight_gb": 5.229167616, "kv_alloc_gb_per_1k_tokens": 0.053248, "kv_read_gb_per_1k_tokens": 0.053248, "kv_alloc_fixed_gb": 0.218103808, "kv_read_fixed_gb": 0.218103808, "notes": "Audited on 2026-07-07 from live Hugging Face metadata, stale repo config-stub check, linked GGUF file metadata, existing gated-base profile evidence, and direct GGUF header range-read at commit bd652eddf75b473fe86639b6b927e06972277d1a. The API-selected artifact is gemma-2-2b-it.fp16.gguf. Legacy adapter fields approximate the Gemma 2 layered KV rule as 13 full-context layers plus 13 sliding-window layers capped at 4096 tokens; the self-contained model profile is authoritative for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/gemma-2-2b-it-GGUF" }, { "label": "Repo config stub", "url": "https://huggingface.co/MaziyarPanahi/gemma-2-2b-it-GGUF/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-2-2b-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/gemma-2-2b-it-GGUF" }, { "label": "Selected FP16 GGUF artifact", "url": "https://huggingface.co/MaziyarPanahi/gemma-2-2b-it-GGUF/resolve/bd652eddf75b473fe86639b6b927e06972277d1a/gemma-2-2b-it.fp16.gguf" } ] }, { "id": "CalamitousFelicitousness/Qwen2.5-32B-Instruct-fp8-dynamic", "name": "Qwen2.5-32B-Instruct-fp8-dynamic FP8", "short_name": "Qwen2.5 32B Instruct fp8 dynamic", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 153510, "tags": [ "safetensors", "qwen2", "chat", "text-generation", "conversational", "en", "arxiv:2309.00071", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-32B", "base_model:quantized:Qwen/Qwen2.5-32B", "license:apache-2.0", "fp8", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM, FP8 dynamic)", "total_params_b": 32.763876352, "active_params_b": 32.763876352, "layers": 64, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "FP8 F8_E4M3 + BF16/F32 side tensors", "weight_bytes_per_param": 1.0475601960909269, "resident_weight_gb": 34.322132736, "active_weight_gb": 32.764997376, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, model card, pinned served FP8 config, existing audited Qwen2.5 32B Instruct profile/config evidence, safetensors index, linked-object metadata, and direct safetensors shard header range reads. Bound weights separate the exact 34.322132736 GB resident FP8/BF16/F32 tensor payload from 32.764997376 GB ordinary text-decode swept traffic; the audited profile supersedes the generated ideal 1-byte dense estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/CalamitousFelicitousness/Qwen2.5-32B-Instruct-fp8-dynamic" }, { "label": "Model config", "url": "https://huggingface.co/CalamitousFelicitousness/Qwen2.5-32B-Instruct-fp8-dynamic/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-32B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/CalamitousFelicitousness/Qwen2.5-32B-Instruct-fp8-dynamic" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/CalamitousFelicitousness/Qwen2.5-32B-Instruct-fp8-dynamic/raw/7ae45d106280752e6961841f0711344dedfcb2cd/model.safetensors.index.json" } ] }, { "id": "inclusionAI/LLaDA2.0-mini", "name": "LLaDA2.0-mini BF16", "short_name": "LLaDA2.0 mini", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 146675, "tags": [ "transformers", "safetensors", "llada2_moe", "text-generation", "dllm", "diffusion", "llm", "text_generation", "conversational", "custom_code", "arxiv:2512.15745", "license:apache-2.0", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "llada2_moe block diffusion (LLaDA2MoeModelLM)", "total_params_b": 16.255643392, "active_params_b": 1.373204224, "layers": 20, "attention_heads": 16, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 32.511286784, "active_weight_gb": 2.746408448, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Reviewed on 2026-07-07 from Hugging Face API metadata, pinned config, model card, custom generation code, and direct safetensors header range reads. Production bounds fail closed: LLaDA2.0 Mini uses block-wise diffusion over masked output tokens, not ordinary autoregressive KV decode." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/inclusionAI/LLaDA2.0-mini" }, { "label": "Pinned model config", "url": "https://huggingface.co/inclusionAI/LLaDA2.0-mini/raw/dad945cac317da394b390f82c7b40691d8a881ed/config.json" }, { "label": "Custom LLaDA2 modeling code", "url": "https://huggingface.co/inclusionAI/LLaDA2.0-mini/raw/dad945cac317da394b390f82c7b40691d8a881ed/modeling_llada2_moe.py" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/inclusionAI/LLaDA2.0-mini/raw/dad945cac317da394b390f82c7b40691d8a881ed/model.safetensors.index.json" } ] }, { "id": "MiniMaxAI/MiniMax-VL-01", "name": "MiniMax-VL-01 F32/BF16", "short_name": "MiniMax VL 01", "license": "other", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 112771, "tags": [ "safetensors", "minimax_vl_01", "image-text-to-text", "conversational", "custom_code", "arxiv:2501.08313", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "minimax_vl_01 (MiniMaxVL01ForConditionalGeneration, MiniMaxText01 Lightning/MoE text)", "total_params_b": 456.437219328, "active_params_b": 47.18984192, "layers": 80, "attention_heads": 64, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 8192, "routed_experts": 32, "routed_experts_per_token": 2, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "F32/BF16", "weight_bytes_per_param": 2.0054601850472866, "resident_weight_gb": 915.366670336, "active_weight_gb": 94.37968384, "expert_param_b": 13.552605781121152, "fixed_param_b": 19.95629503280083, "shared_expert_param_b": 0, "kv_alloc_gb_per_1k_tokens": 0.04096, "kv_read_gb_per_1k_tokens": 0.04096, "kv_alloc_fixed_gb": 0.29360128, "kv_read_fixed_gb": 0.29360128, "notes": "Manual correction on 2026-07-07 from live Hugging Face API metadata, served config, model card, custom text/VL modeling code, safetensors index, and direct shard-header range reads. active_weight_gb records exact single-token ordinary text weight traffic from headers: 40.021504000 GB fixed traffic plus 2 routed expert groups at 27.179089920 GB each. KV fields approximate the profile-backed composite state: 70 Lightning-Attention layers as 0.293601280 GB fixed FP32 recurrent state plus 10 full-attention BF16 K/V layers at 0.040960000 GB per 1k context tokens. The authoritative profile records this as layered_kv." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MiniMaxAI/MiniMax-VL-01" }, { "label": "Model card", "url": "https://huggingface.co/MiniMaxAI/MiniMax-VL-01" }, { "label": "Pinned served config", "url": "https://huggingface.co/MiniMaxAI/MiniMax-VL-01/raw/308b79934be140a43a0fb80f82b4e20d0ebe3cb8/config.json" }, { "label": "Custom text modeling code", "url": "https://huggingface.co/MiniMaxAI/MiniMax-VL-01/raw/308b79934be140a43a0fb80f82b4e20d0ebe3cb8/modeling_minimax_text_01.py" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/MiniMaxAI/MiniMax-VL-01/raw/308b79934be140a43a0fb80f82b4e20d0ebe3cb8/model.safetensors.index.json" } ] }, { "id": "unsloth/Qwen3.6-35B-A3B-NVFP4", "name": "Qwen3.6-35B-A3B-NVFP4 NVFP4", "short_name": "Qwen3.6 35B A3B NVFP4", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 184698, "tags": [ "safetensors", "qwen3_5_moe", "unsloth", "image-text-to-text", "conversational", "base_model:Qwen/Qwen3.6-35B-A3B", "base_model:quantized:Qwen/Qwen3.6-35B-A3B", "license:apache-2.0", "8-bit", "compressed-tensors", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_5_moe_text", "total_params_b": 35.9518, "active_params_b": 3, "layers": 40, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "NVFP4", "weight_bytes_per_param": 0.686036114130586, "resident_weight_gb": 24.664233168, "active_weight_gb": 3.511290096, "expert_param_b": 0.10317217788118828, "fixed_param_b": 4.292851812520491, "shared_expert_param_b": 0, "kv_alloc_gb_per_1k_tokens": 0.02048, "kv_read_gb_per_1k_tokens": 0.02048, "kv_alloc_fixed_gb": 0.06488064, "kv_read_fixed_gb": 0.06488064, "notes": "Audited on 2026-07-07 from current Hugging Face API metadata, pinned config, pinned Unsloth NVFP4 recipe, current base config comparison, model card, and direct single-file safetensors header range reads. Resident tensor payloads sum exactly to 24.664233168 GB. Ordinary text-decode traffic charges lm_head, DeltaNet weights, full-attention weights, shared experts, gates, norms, and eight selected routed experts for 3.511290096 GB. The simple KV coefficient records the BF16 full-attention stream only; the self-contained profile also charges the fixed DeltaNet runtime state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/Qwen3.6-35B-A3B-NVFP4" }, { "label": "Model config", "url": "https://huggingface.co/unsloth/Qwen3.6-35B-A3B-NVFP4/raw/612d523c58522734a1f6dc995bdeebe216647fef/config.json" }, { "label": "Unsloth NVFP4 recipe", "url": "https://huggingface.co/unsloth/Qwen3.6-35B-A3B-NVFP4/raw/612d523c58522734a1f6dc995bdeebe216647fef/nvfp4_recipe.json" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/unsloth/Qwen3.6-35B-A3B-NVFP4/resolve/612d523c58522734a1f6dc995bdeebe216647fef/model.safetensors" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B" } ] }, { "id": "meta-llama/Llama-Guard-4-12B", "name": "Llama-Guard-4-12B BF16", "short_name": "Llama Guard 4 12B", "license": "other", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 153621, "tags": [ "transformers", "safetensors", "llama4", "image-text-to-text", "facebook", "meta", "pytorch", "llama", "safety", "conversational", "en", "arxiv:2503.05731", "arxiv:2407.21783", "license:other", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense multimodal transformer", "detail": "llama4 gated config unavailable", "total_params_b": 12.001097216, "active_params_b": 12.001097216, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 24.002194432, "active_weight_gb": 24.002194432, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manually corrected on 2026-07-07 from current Hugging Face API metadata and raw-file access checks. The repo is manually gated; raw config, README, and safetensors index returned 401 anonymously and 403 with the configured osolmaz HF CLI identity. Production bounds use the unsupported profile and must not infer KV geometry, context length, multimodal residency, or swept decode traffic from this row." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-Guard-4-12B" }, { "label": "Gated config access check", "url": "https://huggingface.co/meta-llama/Llama-Guard-4-12B/raw/87acb4b94e930c3d679e6e7ee9d57e2feab9ea71/config.json" }, { "label": "Hugging Face model card", "url": "https://huggingface.co/meta-llama/Llama-Guard-4-12B" } ] }, { "id": "lmstudio-community/gemma-4-E2B-it-GGUF", "name": "gemma-4-E2B-it-GGUF GGUF", "short_name": "gemma 4 E2B it", "license": "apache-2.0", "hf_pipeline_tag": "unknown", "hf_downloads": 141691, "tags": [ "gguf", "base_model:google/gemma-4-E2B-it", "base_model:quantized:google/gemma-4-E2B-it", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 4.647450147, "active_params_b": 2.298639907, "layers": 35, "attention_heads": 8, "kv_heads": 1, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q4_K_M", "weight_bytes_per_param": 0.7375824565246273, "resident_weight_gb": 3.427877696, "active_weight_gb": 1.4853019, "kv_alloc_gb_per_1k_tokens": 0.006144, "kv_read_gb_per_1k_tokens": 0.014336, "kv_alloc_fixed_gb": 0.006291456, "kv_read_fixed_gb": 0.014680064, "notes": "Audited profile-backed row from live Hugging Face metadata on 2026-07-07; downloads=141691; selected artifact gemma-4-E2B-it-Q4_K_M.gguf; data quality: API metadata, model card, Google base config, GGUF HEAD checks, Transformers PLE documentation, and direct GGUF header tensor index. Resident weight uses linked file size; active weight excludes the resident-only per-layer token embedding lookup table. Legacy adapter fields approximate the shared-KV rule: full-context traffic is token-proportional, while the sliding-window component is capped at 512 tokens; the self-contained model profile is authoritative for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/gemma-4-E2B-it-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-E2B-it-GGUF" }, { "label": "Selected GGUF tensor index", "url": "https://huggingface.co/lmstudio-community/gemma-4-E2B-it-GGUF/resolve/3155b2dec005f7b0a6f721198c26ab11492a38e4/gemma-4-E2B-it-Q4_K_M.gguf" }, { "label": "Pinned Google base config", "url": "https://huggingface.co/google/gemma-4-E2B-it/raw/70af34e20bd4b7a91f0de6b22675850c43922a03/config.json" }, { "label": "Gemma 4 PLE documentation", "url": "https://huggingface.co/docs/transformers/model_doc/gemma4" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-E2B-it" } ] }, { "id": "Qwen/Qwen2.5-Coder-7B-Instruct-GGUF", "name": "Qwen2.5-Coder-7B-Instruct-GGUF GGUF", "short_name": "Qwen2.5 Coder 7B Instruct", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 155704, "tags": [ "transformers", "gguf", "code", "codeqwen", "chat", "qwen", "qwen-coder", "text-generation", "en", "arxiv:2409.12186", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-Coder-7B-Instruct", "base_model:quantized:Qwen/Qwen2.5-Coder-7B-Instruct", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 7.615616512, "active_params_b": 7.070619136, "layers": 28, "attention_heads": 28, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q5_K_M", "weight_bytes_per_param": 0.7149560673676947, "resident_weight_gb": 5.444831232, "active_weight_gb": 5.064192, "kv_alloc_gb_per_1k_tokens": 0.057344, "kv_read_gb_per_1k_tokens": 0.057344, "notes": "Audited profile-backed row from live Hugging Face metadata on 2026-07-07; downloads=155704; selected artifact qwen2.5-coder-7b-instruct-q5_k_m.gguf because the pinned model-card CLI example selects Q5_K_M. Data quality: API metadata, pinned model card, audited BF16 base profile, GGUF HEAD checks, and direct GGUF header tensor index. Resident weight uses linked file size; active weight excludes the resident-only input embedding lookup table. The selected GGUF header and HF API record 131072 context, while the card text says GGUF full context is 32768; the self-contained profile records this mismatch and is authoritative for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Coder-7B-Instruct-GGUF" }, { "label": "Selected GGUF tensor index", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct-GGUF/resolve/13fb94bfda8c8cf22497dc57b78f391a9acb426a/qwen2.5-coder-7b-instruct-q5_k_m.gguf" }, { "label": "Pinned base config", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct/raw/c03e6d358207e414f1eca0bb1891e29f1db0e242/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct" } ] }, { "id": "MaziyarPanahi/Yi-Coder-9B-Chat-GGUF", "name": "Yi-Coder-9B-Chat-GGUF GGUF Q2_K", "short_name": "Yi Coder 9B Chat", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 118204, "tags": [ "gguf", "mistral", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:01-ai/Yi-Coder-9B-Chat", "base_model:quantized:01-ai/Yi-Coder-9B-Chat", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "llama GGUF (LlamaForCausalLM)", "total_params_b": 8.829407232, "active_params_b": 8.567263232, "layers": 48, "attention_heads": 32, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q2_K", "weight_bytes_per_param": 0.3799038490197934, "resident_weight_gb": 3.354325792, "active_weight_gb": 3.26680576, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "notes": "Audited profile-backed row from live Hugging Face metadata on 2026-07-07; downloads=118204; selected artifact Yi-Coder-9B-Chat.Q2_K.gguf because this is the catalog's 2-bit GGUF row. Data quality: API metadata, model card, pinned base config, stale repo config check, GGUF HEAD checks, and direct GGUF header tensor index. Resident weight uses linked file size; active weight excludes the resident-only input embedding lookup table. The repo-local config.json says model_type mistral but conflicts with the base config and selected GGUF header, so it is not used as architecture evidence." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/Yi-Coder-9B-Chat-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Yi-Coder-9B-Chat-GGUF" }, { "label": "Selected GGUF tensor index", "url": "https://huggingface.co/MaziyarPanahi/Yi-Coder-9B-Chat-GGUF/resolve/a1ceaae2d11b41e263560af6f806b9aedbefffd2/Yi-Coder-9B-Chat.Q2_K.gguf" }, { "label": "Pinned base config", "url": "https://huggingface.co/01-ai/Yi-Coder-9B-Chat/raw/356a1f8d4e4a606d0b879e54191ca809918576b8/config.json" }, { "label": "Stale repo config check", "url": "https://huggingface.co/MaziyarPanahi/Yi-Coder-9B-Chat-GGUF/raw/a1ceaae2d11b41e263560af6f806b9aedbefffd2/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/01-ai/Yi-Coder-9B-Chat" } ] }, { "id": "TheBloke/Mistral-7B-Instruct-v0.2-GGUF", "name": "Mistral-7B-Instruct-v0.2-GGUF GGUF Q4_K_M", "short_name": "Mistral 7B Instruct v0.2", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 150845, "tags": [ "transformers", "gguf", "mistral", "finetuned", "text-generation", "arxiv:2310.06825", "base_model:mistralai/Mistral-7B-Instruct-v0.2", "base_model:quantized:mistralai/Mistral-7B-Instruct-v0.2", "license:apache-2.0", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "mistral GGUF (MistralForCausalLM)", "total_params_b": 7.241732096, "active_params_b": 7.110660096, "layers": 32, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q4_K_M", "weight_bytes_per_param": 0.603231316222389, "resident_weight_gb": 4.368439584, "active_weight_gb": 4.293976064, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "notes": "Audited on 2026-07-07 from live Hugging Face metadata, pinned base config, stale repo-config check, linked-object size checks, and direct selected Q4_K_M GGUF tensor-index range read. The catalog snapshot downloads count is retained for threshold validation; the live API count during audit was 86403. The card examples select Q4_K_M, while API gguf.totalFileSize points at Q2_K, so Q4_K_M is the row's selected artifact." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/TheBloke/Mistral-7B-Instruct-v0.2-GGUF" }, { "label": "Pinned model card", "url": "https://huggingface.co/TheBloke/Mistral-7B-Instruct-v0.2-GGUF" }, { "label": "Selected GGUF tensor index", "url": "https://huggingface.co/TheBloke/Mistral-7B-Instruct-v0.2-GGUF/resolve/3a6fbf4a41a1d52e415a4958cde6856d34b2db93/mistral-7b-instruct-v0.2.Q4_K_M.gguf" }, { "label": "Pinned base config", "url": "https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2/raw/63a8b081895390a26e140280378bc85ec8bce07a/config.json" }, { "label": "Stale repo config check", "url": "https://huggingface.co/TheBloke/Mistral-7B-Instruct-v0.2-GGUF/raw/3a6fbf4a41a1d52e415a4958cde6856d34b2db93/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2" } ] }, { "id": "MaziyarPanahi/Qwen2.5-1.5B-Instruct-GGUF", "name": "Qwen2.5-1.5B-Instruct-GGUF GGUF Q2_K", "short_name": "Qwen2.5 1.5B Instruct", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 149795, "tags": [ "gguf", "mistral", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:Qwen/Qwen2.5-1.5B-Instruct", "base_model:quantized:Qwen/Qwen2.5-1.5B-Instruct", "license:apache-2.0", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 GGUF Q2_K", "total_params_b": 1.543714304, "active_params_b": 1.543714304, "layers": 28, "attention_heads": 12, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q2_K", "weight_bytes_per_param": 0.43810251563232255, "resident_weight_gb": 0.67630512, "active_weight_gb": 0.67035392, "kv_alloc_gb_per_1k_tokens": 0.028672, "kv_read_gb_per_1k_tokens": 0.028672, "notes": "Audited on 2026-07-07 from current Hugging Face metadata, pinned base config, stale repo-config check, linked-object size checks, and direct selected Q2_K GGUF tensor-index range read. The catalog snapshot downloads count is retained for threshold validation; the live API count during audit was 117624. The API gguf.totalFileSize points at the FP16 sibling, so Q2_K is selected from the catalog's 2-bit row." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Qwen2.5-1.5B-Instruct-GGUF" }, { "label": "Pinned model card", "url": "https://huggingface.co/MaziyarPanahi/Qwen2.5-1.5B-Instruct-GGUF" }, { "label": "Selected GGUF tensor index", "url": "https://huggingface.co/MaziyarPanahi/Qwen2.5-1.5B-Instruct-GGUF/resolve/f67db4478e8de736c5b4f05eac67a9a84f1376a3/Qwen2.5-1.5B-Instruct.Q2_K.gguf" }, { "label": "Pinned base config", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct/raw/989aa7980e4cf806f80c7fef2b1adb7bc71aa306/config.json" }, { "label": "Stale repo config check", "url": "https://huggingface.co/MaziyarPanahi/Qwen2.5-1.5B-Instruct-GGUF/raw/f67db4478e8de736c5b4f05eac67a9a84f1376a3/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct" } ] }, { "id": "nvidia/NVLM-D-72B", "name": "NVLM-D-72B F32/BF16", "short_name": "NVLM D 72B", "license": "cc-by-nc-4.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 149066, "tags": [ "transformers", "safetensors", "NVLM_D", "nvidia", "NVLM", "pytorch", "multimodal", "conversational", "image-text-to-text", "custom_code", "en", "arxiv:2409.11402", "license:cc-by-nc-4.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "NVLM_D multimodal / Qwen2 text decoder", "total_params_b": 79.379593344, "active_params_b": 71.46049536, "layers": 80, "attention_heads": 64, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16/F32", "weight_bytes_per_param": 2.2285591894301557, "resident_weight_gb": 176.9021222, "active_weight_gb": 156.34440192, "kv_alloc_gb_per_1k_tokens": 0.32768, "kv_read_gb_per_1k_tokens": 0.32768, "notes": "Audited on 2026-07-07 from current Hugging Face metadata, served NVLM_D config, model card, safetensors index, and direct safetensors header grouping across all 46 shards. Resident weight includes the loaded language, vision, projector, and header/container bytes; ordinary text-decode traffic sweeps only language layers, final norm, and lm_head. Multimodal image prefill requires a separate workload adapter." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/NVLM-D-72B" }, { "label": "Pinned model card", "url": "https://huggingface.co/nvidia/NVLM-D-72B" }, { "label": "Pinned model config", "url": "https://huggingface.co/nvidia/NVLM-D-72B/raw/11e9e0f8d265d65ddaa8f9445b6f55fdec9d686f/config.json" }, { "label": "Safetensors index", "url": "https://huggingface.co/nvidia/NVLM-D-72B/raw/11e9e0f8d265d65ddaa8f9445b6f55fdec9d686f/model.safetensors.index.json" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/nvidia/NVLM-D-72B/tree/11e9e0f8d265d65ddaa8f9445b6f55fdec9d686f" } ] }, { "id": "MaziyarPanahi/Llama-3.2-3B-Instruct-GGUF", "name": "Llama-3.2-3B-Instruct-GGUF GGUF Q2_K", "short_name": "Llama 3.2 3B Instruct", "license": "llama3.2", "hf_pipeline_tag": "text-generation", "hf_downloads": 148669, "tags": [ "gguf", "mistral", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:meta-llama/Llama-3.2-3B-Instruct", "base_model:quantized:meta-llama/Llama-3.2-3B-Instruct", "license:llama3.2", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "llama GGUF Q2_K (Llama 3.2 3B Instruct)", "total_params_b": 3.212749888, "active_params_b": 3.212749888, "layers": 28, "attention_heads": 24, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q2_K", "weight_bytes_per_param": 0.4245384245734351, "resident_weight_gb": 1.363935776, "active_weight_gb": 1.356097792, "kv_alloc_gb_per_1k_tokens": 0.114688, "kv_read_gb_per_1k_tokens": 0.114688, "notes": "Manually audited on 2026-07-07 from current Hugging Face 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. The authoritative bounds profile uses Llama-3.2-3B-Instruct.Q2_K.gguf for the catalog's 2-bit row: 1.363935776 GB resident and 1.356097792 GB ordinary text-decode swept traffic. The live API gguf.totalFileSize currently points to the FP16 sibling, so selected-file header bytes are authoritative." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/Llama-3.2-3B-Instruct-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/MaziyarPanahi/Llama-3.2-3B-Instruct-GGUF/raw/e56a0ae870579697698c3ded68df97747125d554/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Llama-3.2-3B-Instruct-GGUF" }, { "label": "Selected Q2_K GGUF tensor-index audit", "url": "https://huggingface.co/MaziyarPanahi/Llama-3.2-3B-Instruct-GGUF/resolve/e56a0ae870579697698c3ded68df97747125d554/Llama-3.2-3B-Instruct.Q2_K.gguf" }, { "label": "Base model card", "url": "https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct" }, { "label": "Base model API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-3.2-3B-Instruct" } ] }, { "id": "mratsim/GLM-4-32B-0414.w4a16-gptq", "name": "GLM-4-32B-0414.w4a16-gptq GPTQ W4A16", "short_name": "GLM 4 32B 0414.w4a16", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 148059, "tags": [ "safetensors", "glm4", "gptq", "vllm", "llmcompressor", "text-generation-inference", "text-generation", "conversational", "dataset:mit-han-lab/pile-val-backup", "base_model:zai-org/GLM-4-32B-0414", "base_model:quantized:zai-org/GLM-4-32B-0414", "license:mit", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "glm4 GPTQ W4A16 (Glm4ForCausalLM)", "total_params_b": 32.566081536, "active_params_b": 31.634946048, "layers": 61, "attention_heads": 48, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GPTQ W4A16", "weight_bytes_per_param": 0.6042595387549667, "resident_weight_gb": 19.678365408, "active_weight_gb": 17.816094432, "kv_alloc_gb_per_1k_tokens": 0.062464, "kv_read_gb_per_1k_tokens": 0.062464, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, README, served GPTQ config, recipe.yaml, base API/config comparison, safetensors index, direct shard header range reads, linked-object HEAD checks, and the Transformers GLM4 implementation. The authoritative bounds profile uses exact compressed-tensors GPTQ payload bytes: 19.678365408 GB resident and 17.816094432 GB ordinary text-decode swept traffic after excluding model.embed_tokens.weight. The live API downloads were 74733 during audit; catalog snapshot downloads remain 148059." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/mratsim/GLM-4-32B-0414.w4a16-gptq" }, { "label": "Model config", "url": "https://huggingface.co/mratsim/GLM-4-32B-0414.w4a16-gptq/raw/7649a2140aa2dd535d23c8794be7662f23a7aee8/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/mratsim/GLM-4-32B-0414.w4a16-gptq" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/mratsim/GLM-4-32B-0414.w4a16-gptq/resolve/7649a2140aa2dd535d23c8794be7662f23a7aee8/model.safetensors.index.json" }, { "label": "Quantization recipe", "url": "https://huggingface.co/mratsim/GLM-4-32B-0414.w4a16-gptq/raw/7649a2140aa2dd535d23c8794be7662f23a7aee8/recipe.yaml" }, { "label": "Base model card", "url": "https://huggingface.co/zai-org/GLM-4-32B-0414" }, { "label": "Base model config", "url": "https://huggingface.co/zai-org/GLM-4-32B-0414/raw/077b5c2f5c43bd3239fd605a0600229e8facbd4a/config.json" } ] }, { "id": "stepfun-ai/Step-3.7-Flash", "name": "Step-3.7-Flash BF16/F32", "short_name": "Step 3.7 Flash", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 147982, "tags": [ "transformers", "safetensors", "step3p7", "text-generation", "vision-language", "multimodal", "moe", "image-text-to-text", "conversational", "custom_code", "en", "license:apache-2.0", "eval-results" ], "architecture": { "type": "MoE transformer", "detail": "step3p7 VLM with step3p5 MoE language model", "total_params_b": 201.36531616, "active_params_b": 11.459378048, "layers": 45, "attention_heads": 64, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 288, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "BF16/F32", "weight_bytes_per_param": 2.000000120139856, "resident_weight_gb": 402.730656512, "active_weight_gb": 22.918756096, "kv_alloc_gb_per_1k_tokens": 0.049152, "kv_read_gb_per_1k_tokens": 0.049152, "kv_alloc_fixed_gb": 0.069206016, "kv_read_fixed_gb": 0.069206016, "expert_param_b": 0.66060288, "shared_expert_param_b": 0, "fixed_param_b": 6.174555008, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, model card, pinned served config, remote modeling code, safetensors index, linked shard HEAD checks, and direct shard-header range reads. Live API downloads during audit were 146322; catalog snapshot downloads are retained at 147982. Bound weights use exact stored bytes: 402.730656512 GB resident package, 12.349110016 GB fixed ordinary traffic, 1.321205760 GB per routed expert index across 288 experts, and 9.874287616 GB resident-only input embedding, vision/projector, and MTP tensors. KV traffic uses BF16 full/sliding cache: 12 full-context layers plus 33 sliding-window layers capped at 512 tokens." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/stepfun-ai/Step-3.7-Flash" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/stepfun-ai/Step-3.7-Flash" }, { "label": "Pinned served config", "url": "https://huggingface.co/stepfun-ai/Step-3.7-Flash/raw/5f6244077ac62e04eec3f320501ff8c2b293373a/config.json" }, { "label": "Custom modeling code", "url": "https://huggingface.co/stepfun-ai/Step-3.7-Flash/raw/5f6244077ac62e04eec3f320501ff8c2b293373a/modeling_step3p7.py" }, { "label": "Safetensors index and shard-header audit", "url": "https://huggingface.co/stepfun-ai/Step-3.7-Flash/resolve/5f6244077ac62e04eec3f320501ff8c2b293373a/model.safetensors.index.json" } ] }, { "id": "MaziyarPanahi/gemma-3-12b-it-GGUF", "name": "gemma-3-12b-it-GGUF FP16", "short_name": "Gemma 3 12B IT", "license": "gemma", "hf_pipeline_tag": "text-generation", "hf_downloads": 117347, "tags": [ "gguf", "mistral", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:google/gemma-3-12b-it", "base_model:quantized:google/gemma-3-12b-it", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "gemma3 (GGUF FP16; package config says mistral)", "total_params_b": 11.765788416, "active_params_b": 11.765788416, "layers": 48, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF FP16/F32", "weight_bytes_per_param": 2, "resident_weight_gb": 23.539658528, "active_weight_gb": 23.533108224, "kv_alloc_gb_per_1k_tokens": 0.065536, "kv_read_gb_per_1k_tokens": 0.065536, "kv_alloc_fixed_gb": 0.33554432, "kv_read_fixed_gb": 0.33554432, "notes": "Audited 2026-07-07 from live HF API metadata, HF CLI dry-run, exact linked-object HEAD checks, package config conflict check, gated-base access check, Google Gemma 3 local/global attention documentation, and a range-read of the selected FP16 GGUF header; downloads=117347. API gguf.totalFileSize selects gemma-3-12b-it.fp16.gguf, so this row intentionally models the FP16 artifact rather than lower-bit GGUF siblings advertised by tags. Exact selected file residency is 23.539658528 GB and ordinary swept tensor traffic is 23.533108224 GB. KV uses Gemma 3 layered FP16 K/V over eight full-context layers plus 40 sliding-window layers capped at 1024 tokens." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/gemma-3-12b-it-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/gemma-3-12b-it-GGUF" }, { "label": "Package config conflict check", "url": "https://huggingface.co/MaziyarPanahi/gemma-3-12b-it-GGUF/raw/0e7c92b7b0890410acb9638a0107c8722a6ab1df/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-3-12b-it" }, { "label": "Selected FP16 GGUF artifact", "url": "https://huggingface.co/MaziyarPanahi/gemma-3-12b-it-GGUF/resolve/0e7c92b7b0890410acb9638a0107c8722a6ab1df/gemma-3-12b-it.fp16.gguf" } ] }, { "id": "Qwen/Qwen2.5-Math-7B-Instruct", "name": "Qwen2.5-Math-7B-Instruct BF16", "short_name": "Qwen2.5 Math 7B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 148342, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "chat", "conversational", "en", "arxiv:2409.12122", "base_model:Qwen/Qwen2.5-Math-7B", "base_model:finetune:Qwen/Qwen2.5-Math-7B", "license:apache-2.0", "eval-results", "text-generation-inference", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 7.615616512, "active_params_b": 7.070619136, "layers": 28, "attention_heads": 28, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 4096, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 15.231233024, "active_weight_gb": 14.141238272, "kv_alloc_gb_per_1k_tokens": 0.057344, "kv_read_gb_per_1k_tokens": 0.057344, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, pinned served config, safetensors index, linked-object HEAD checks, and direct shard-header range reads. Exact BF16 payload bytes are 15.231233024 GB resident and 14.141238272 GB swept ordinary text traffic after excluding the resident-only input embedding. The config records separate input and lm_head embeddings, 28 layers, 4 KV heads, 128 head dimension, 4096 context, and use_sliding_window false." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Math-7B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Math-7B-Instruct" }, { "label": "Pinned model config", "url": "https://huggingface.co/Qwen/Qwen2.5-Math-7B-Instruct/raw/ef9926d75ab1d54532f6a30dd5e760355eb9aa4d/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Qwen/Qwen2.5-Math-7B-Instruct/resolve/ef9926d75ab1d54532f6a30dd5e760355eb9aa4d/model.safetensors.index.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Math-7B" } ] }, { "id": "ReadyArt/Melody1437-26B-A4B-v2.0-GGUF", "name": "Melody1437-26B-A4B-v2.0-GGUF HB16 Q4_K_M", "short_name": "Melody1437 26B A4B v2.0", "license": "apache-2.0", "hf_pipeline_tag": "unknown", "hf_downloads": 156079, "tags": [ "gguf", "gemma-4", "roleplay", "conversational", "instruct", "apache-2.0", "nsfw", "explicit", "erp", "adult-content", "mature", "unaligned", "base_model:ReadyArt/Melody1437-26B-A4B-v2.0", "base_model:quantized:ReadyArt/Melody1437-26B-A4B-v2.0", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "gemma4_text (GGUF HB16 Q4_K_M)", "total_params_b": 25.233142046, "active_params_b": 4, "layers": 30, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "GGUF HB16 Q4_K_M", "weight_bytes_per_param": 0.7001450089645328, "resident_weight_gb": 17.666858464, "active_weight_gb": 3.466505336, "kv_alloc_gb_per_1k_tokens": 0.20992, "kv_read_gb_per_1k_tokens": 0.20992, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, HF CLI dry-run, model card metadata, gated base access check, linked GGUF file sizes, and a direct GGUF tensor-index range read. The authoritative bounds profile uses the API-selected Melody1437-26B-A4B-v2.0-HB16-Q4_K_M.gguf artifact: 17.666858464 GB resident, 2.520870008 GB fixed traffic, 0.118204416 GB per routed expert, 8 routed experts per token, 128 routed experts, 1 shared expert charged in fixed traffic, and layered Gemma 4 hybrid KV. Legacy per-1k KV fields show the short-context unsaturated coefficient; long-context calculation must use the profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/ReadyArt/Melody1437-26B-A4B-v2.0-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/ReadyArt/Melody1437-26B-A4B-v2.0-GGUF" }, { "label": "Base model card", "url": "https://huggingface.co/ReadyArt/Melody1437-26B-A4B-v2.0" }, { "label": "Selected HB16 Q4_K_M GGUF artifact", "url": "https://huggingface.co/ReadyArt/Melody1437-26B-A4B-v2.0-GGUF/resolve/b22520a5c83db6ff0d12aef1fd461c932ba86fbf/Melody1437-26B-A4B-v2.0-HB16-Q4_K_M.gguf" } ] }, { "id": "moonshotai/Kimi-K2-Thinking", "name": "Kimi-K2-Thinking INT4/BF16", "short_name": "Kimi K2 Thinking", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 139650, "tags": [ "transformers", "safetensors", "kimi_k2", "text-generation", "conversational", "custom_code", "license:other", "eval-results", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "kimi_k2 (DeepseekV3ForCausalLM)", "total_params_b": 1058.118284416, "active_params_b": 32, "layers": 61, "attention_heads": 64, "kv_heads": 64, "head_dim": 192, "max_context_tokens": 262144, "routed_experts": 384, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "INT4/BF16 compressed-tensors", "weight_bytes_per_param": 0.561568521462566, "resident_weight_gb": 594.205920512, "active_weight_gb": 32.986532352, "kv_alloc_gb_per_1k_tokens": 2.49856, "kv_read_gb_per_1k_tokens": 2.49856, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, served Kimi K2 config, model card, deployment guide, byte-identical Kimi K2 custom runtime review, safetensors index, and direct shard-header range reads. Bound weights separate the exact 594.205920512 GB resident compressed-tensors package from 32.986532352 GB single-session ordinary text-decode MoE traffic; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/moonshotai/Kimi-K2-Thinking" }, { "label": "Model config", "url": "https://huggingface.co/moonshotai/Kimi-K2-Thinking/raw/a51ccc050d73dab088bf7b0e2dd9b30ae85a4e55/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/moonshotai/Kimi-K2-Thinking" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/moonshotai/Kimi-K2-Thinking/resolve/a51ccc050d73dab088bf7b0e2dd9b30ae85a4e55/model.safetensors.index.json" }, { "label": "Custom DeepSeek runtime", "url": "https://huggingface.co/moonshotai/Kimi-K2-Thinking/raw/a51ccc050d73dab088bf7b0e2dd9b30ae85a4e55/modeling_deepseek.py" }, { "label": "Deployment guide", "url": "https://huggingface.co/moonshotai/Kimi-K2-Thinking/raw/a51ccc050d73dab088bf7b0e2dd9b30ae85a4e55/docs/deploy_guidance.md" } ] }, { "id": "unsloth/Qwen3.5-2B-GGUF", "name": "Qwen3.5-2B-GGUF BF16", "short_name": "Qwen3.5 2B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 140054, "tags": [ "transformers", "gguf", "unsloth", "image-text-to-text", "base_model:Qwen/Qwen3.5-2B", "base_model:quantized:Qwen/Qwen3.5-2B", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text (selected BF16 GGUF)", "total_params_b": 1.881825088, "active_params_b": 1.881825088, "layers": 24, "attention_heads": 8, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16/F32 GGUF", "weight_bytes_per_param": 2.000583127521786, "resident_weight_gb": 3.775709216, "active_weight_gb": 3.76474752, "kv_alloc_gb_per_1k_tokens": 0.012288, "kv_read_gb_per_1k_tokens": 0.012288, "kv_alloc_fixed_gb": 0.019759104, "kv_read_fixed_gb": 0.019759104, "notes": "Manually audited on 2026-07-07 from HF API commit f6d5376be1edb4d416d56da11e5397a961aca8ae, Qwen base config, model card, linked GGUF file sizes, and a GGUF tensor-index range read. The authoritative bounds profile uses the selected Qwen3.5-2B-BF16.gguf artifact: 3.775709216 GB resident, 3.764747520 GB swept text traffic including tied token_embd.weight output projection, 0.010961696 GB GGUF metadata/header/file overhead, and Qwen3.5 hybrid KV plus fixed DeltaNet state. Separate mmproj GGUF files are not included unless explicitly loaded." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/Qwen3.5-2B-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3.5-2B/raw/15852e8c16360a2fea060d615a32b45270f8a8fc/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-2B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3.5-2B-GGUF" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/unsloth/Qwen3.5-2B-GGUF/resolve/f6d5376be1edb4d416d56da11e5397a961aca8ae/Qwen3.5-2B-BF16.gguf" } ] }, { "id": "nm-testing/Qwen1.5-MoE-A2.7B-Chat-quantized.w4a16", "name": "Qwen1.5-MoE-A2.7B-Chat-quantized.w4a16 W4A16", "short_name": "Qwen1.5 MoE A2.7B Chat quantized.w4a16", "license": "unknown", "hf_pipeline_tag": "unknown", "hf_downloads": 143934, "tags": [ "safetensors", "qwen2_moe", "compressed-tensors", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen2_moe (Qwen2MoeForCausalLM, compressed-tensors W4A16)", "total_params_b": 14.42274792, "active_params_b": 2.7, "layers": 24, "attention_heads": 16, "kv_heads": 16, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": 60, "routed_experts_per_token": 4, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "compressed-tensors GPTQ W4A16 packed I32 + BF16/I64 side tensors", "weight_bytes_per_param": 0.5761895970237549, "resident_weight_gb": 8.310237312, "active_weight_gb": 1.692871808, "kv_alloc_gb_per_1k_tokens": 0.196608, "kv_read_gb_per_1k_tokens": 0.196608, "expert_param_b": 0.185796842832597, "fixed_param_b": 2.1948590924453315, "shared_expert_param_b": 0, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, pinned compressed-tensors config, public base chat model card/config comparison, recipe.yaml, safetensors index, linked-object metadata, and direct range-read safetensors shard headers. Legacy fallback values mirror the audited profile: exact 8.310237312 GB resident payload, 1.264654976 GB fixed always-on ordinary text traffic including shared experts, 0.107054208 GB per routed expert index across 60 experts, and 0.622329856 GB resident-only input embeddings. The authoritative profile supersedes the generated ideal flat 4-bit estimate." }, "sources": [ { "label": "Hugging Face model repo", "url": "https://huggingface.co/nm-testing/Qwen1.5-MoE-A2.7B-Chat-quantized.w4a16" }, { "label": "Model config", "url": "https://huggingface.co/nm-testing/Qwen1.5-MoE-A2.7B-Chat-quantized.w4a16/raw/44f1710adc798a746a42b8f43165bb8efe9a71f4/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen1.5-MoE-A2.7B-Chat" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nm-testing/Qwen1.5-MoE-A2.7B-Chat-quantized.w4a16" }, { "label": "Pinned base config", "url": "https://huggingface.co/Qwen/Qwen1.5-MoE-A2.7B-Chat/raw/ec052fda178e241c7c443468d2fa1db6618996be/config.json" }, { "label": "Quantization recipe", "url": "https://huggingface.co/nm-testing/Qwen1.5-MoE-A2.7B-Chat-quantized.w4a16/raw/44f1710adc798a746a42b8f43165bb8efe9a71f4/recipe.yaml" }, { "label": "Safetensors index and shard-header audit", "url": "https://huggingface.co/nm-testing/Qwen1.5-MoE-A2.7B-Chat-quantized.w4a16/resolve/44f1710adc798a746a42b8f43165bb8efe9a71f4/model.safetensors.index.json" } ] }, { "id": "MaziyarPanahi/Mistral-Small-24B-Instruct-2501-GGUF", "name": "Mistral-Small-24B-Instruct-2501-GGUF GGUF F16", "short_name": "Mistral Small 24B Instruct 2501", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 116025, "tags": [ "gguf", "mistral", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:mistralai/Mistral-Small-24B-Instruct-2501", "base_model:quantized:mistralai/Mistral-Small-24B-Instruct-2501", "license:apache-2.0", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "llama GGUF / MistralForCausalLM", "total_params_b": 23.5724032, "active_params_b": 22.90131456, "layers": 40, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF F16/F32", "weight_bytes_per_param": 2, "resident_weight_gb": 47.153518016, "active_weight_gb": 45.80345856, "kv_alloc_gb_per_1k_tokens": 0.16384, "kv_read_gb_per_1k_tokens": 0.16384, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, model card, base Mistral Small config/API metadata, linked GGUF file sizes, and direct selected F16 GGUF tensor-index range read. The authoritative bounds profile uses Mistral-Small-24B-Instruct-2501.fp16.gguf because it matches API gguf.totalFileSize: 47.153518016 GB resident and 45.803458560 GB ordinary text-decode swept traffic after excluding token_embd.weight plus GGUF metadata/header/file overhead. The audited profile supersedes the generated 2-bit, 2k-context, zero-KV placeholder." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/Mistral-Small-24B-Instruct-2501-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/MaziyarPanahi/Mistral-Small-24B-Instruct-2501-GGUF/raw/458f84af662e5b23b03e8191e9c3c4d6ab857283/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/mistralai/Mistral-Small-24B-Instruct-2501" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Mistral-Small-24B-Instruct-2501-GGUF" }, { "label": "Pinned base config", "url": "https://huggingface.co/mistralai/Mistral-Small-24B-Instruct-2501/raw/9527884be6e5616bdd54de542f9ae13384489724/config.json" }, { "label": "Selected F16 GGUF tensor-index audit", "url": "https://huggingface.co/MaziyarPanahi/Mistral-Small-24B-Instruct-2501-GGUF/resolve/458f84af662e5b23b03e8191e9c3c4d6ab857283/Mistral-Small-24B-Instruct-2501.fp16.gguf" } ] }, { "id": "nvidia/Qwen3-8B-NVFP4", "name": "Qwen3-8B-NVFP4 NVFP4", "short_name": "Qwen3 8B NVFP4", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 150764, "tags": [ "Model Optimizer", "safetensors", "qwen3", "nvidia", "ModelOpt", "Qwen3", "quantized", "FP4", "fp4", "text-generation", "conversational", "base_model:Qwen/Qwen3-8B", "base_model:quantized:Qwen/Qwen3-8B", "license:apache-2.0", "8-bit", "modelopt", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 8.19073536, "active_params_b": 7.568405504, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "ModelOpt NVFP4 + FP8 scales + BF16 fixed tensors", "weight_bytes_per_param": 0.5, "resident_weight_gb": 6.396932352, "active_weight_gb": 5.15227264, "kv_alloc_gb_per_1k_tokens": 0.073728, "kv_read_gb_per_1k_tokens": 0.073728, "notes": "Manual correction on 2026-07-07 from current Hugging Face API metadata, NVIDIA model card, pinned served config, ModelOpt quantization config, current Qwen3 8B base config comparison, safetensors index, linked-object HEAD checks, and direct safetensors shard-header range reads. Legacy fallback values mirror the audited profile: exact 6.396932352 GB stored resident package, 5.152272640 GB ordinary text-decode swept traffic, 1.244659712 GB resident-only input embedding, and FP8 full-context KV. The authoritative profile supersedes the generated simple 0.5-byte NVFP4 resident estimate and BF16 KV estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/Qwen3-8B-NVFP4" }, { "label": "Model config", "url": "https://huggingface.co/nvidia/Qwen3-8B-NVFP4/raw/ccd10a893cbca613259517c3efe08e151ddf2b8e/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-8B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/Qwen3-8B-NVFP4" }, { "label": "ModelOpt quantization config", "url": "https://huggingface.co/nvidia/Qwen3-8B-NVFP4/raw/ccd10a893cbca613259517c3efe08e151ddf2b8e/hf_quant_config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/nvidia/Qwen3-8B-NVFP4/raw/ccd10a893cbca613259517c3efe08e151ddf2b8e/model.safetensors.index.json" } ] }, { "id": "unsloth/Qwen3.5-27B-GGUF", "name": "Qwen3.5-27B-GGUF GGUF", "short_name": "Qwen3.5 27B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 131531, "tags": [ "gguf", "unsloth", "image-text-to-text", "base_model:Qwen/Qwen3.5-27B", "base_model:quantized:Qwen/Qwen3.5-27B", "license:apache-2.0", "endpoints_compatible", "region:us", "imatrix", "conversational" ], "architecture": { "type": "Dense multimodal transformer", "detail": "qwen3_5_text (API-selected IQ4_NL GGUF)", "total_params_b": 26.895998464, "active_params_b": 25.624600064, "layers": 64, "attention_heads": 24, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF IQ4_NL", "weight_bytes_per_param": 0.5832798869689881, "resident_weight_gb": 15.687894944, "active_weight_gb": 14.961739776, "kv_alloc_gb_per_1k_tokens": 0.065536, "kv_read_gb_per_1k_tokens": 0.065536, "kv_alloc_fixed_gb": 0.154927104, "kv_read_fixed_gb": 0.154927104, "notes": "Manually audited on 2026-07-07 from HF API commit 3221f178a6b842d04f1fb42f1c413534adcc0a6a, model card, pinned Qwen base config, linked GGUF file sizes, and a GGUF tensor-index range read. The authoritative bounds profile uses the API-selected Qwen3.5-27B-IQ4_NL.gguf artifact: 15.687894944 GB resident, 14.961739776 GB swept ordinary text traffic, 0.726155168 GB resident-only token embedding plus GGUF metadata/header/file overhead, and Qwen3.5 hybrid KV plus fixed DeltaNet state. Separate mmproj sidecars are not included unless explicitly loaded." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/Qwen3.5-27B-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3.5-27B/raw/fc05daec18b0a78c049392ed2e771dde82bdf654/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-27B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3.5-27B-GGUF" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/unsloth/Qwen3.5-27B-GGUF/resolve/3221f178a6b842d04f1fb42f1c413534adcc0a6a/Qwen3.5-27B-IQ4_NL.gguf" } ] }, { "id": "cortecs/Llama-3.3-70B-Instruct-FP8-Dynamic", "name": "Llama-3.3-70B-Instruct-FP8-Dynamic FP8", "short_name": "Llama 3.3 70B Instruct FP8 Dynamic", "license": "other", "hf_pipeline_tag": "unknown", "hf_downloads": 145566, "tags": [ "safetensors", "llama", "base_model:meta-llama/Llama-3.3-70B-Instruct", "base_model:quantized:meta-llama/Llama-3.3-70B-Instruct", "license:other", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "llama (LlamaForCausalLM)", "total_params_b": 70.560423936, "active_params_b": 69.509750784, "layers": 80, "attention_heads": 64, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "FP8", "weight_bytes_per_param": 1, "resident_weight_gb": 72.669806592, "active_weight_gb": 70.568460288, "kv_alloc_gb_per_1k_tokens": 0.32768, "kv_read_gb_per_1k_tokens": 0.32768, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, pinned model card, served FP8 dynamic config, gated Llama 3.3 base API metadata, safetensors index, linked-object HEAD checks, direct safetensors shard header range reads, and audited public Llama 3.3 derivative config comparison. Bound weights separate the 72.669806592 GB resident compressed-tensors package from the 70.568460288 GB ordinary text-decode swept subset; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/cortecs/Llama-3.3-70B-Instruct-FP8-Dynamic" }, { "label": "Model config", "url": "https://huggingface.co/cortecs/Llama-3.3-70B-Instruct-FP8-Dynamic/raw/3722358cc2b990b22304489b2f87ef3bb876d6f6/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/cortecs/Llama-3.3-70B-Instruct-FP8-Dynamic" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/cortecs/Llama-3.3-70B-Instruct-FP8-Dynamic/raw/3722358cc2b990b22304489b2f87ef3bb876d6f6/model.safetensors.index.json" } ] }, { "id": "ReadyArt/Serenity-26B-A4B-GGUF", "name": "Serenity-26B-A4B-GGUF HB16 Q6_K", "short_name": "Serenity 26B A4B", "license": "apache-2.0", "hf_pipeline_tag": "unknown", "hf_downloads": 151391, "tags": [ "gguf", "gemma-4", "roleplay", "conversational", "instruct", "apache-2.0", "nsfw", "adult-content", "unaligned", "mature", "explicit", "erp", "base_model:ReadyArt/Serenity-26B-A4B", "base_model:quantized:ReadyArt/Serenity-26B-A4B", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "gemma4_text (GGUF HB16 Q6_K)", "total_params_b": 25.233142046, "active_params_b": 4, "layers": 30, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "GGUF HB16 Q6_K", "weight_bytes_per_param": 0.9316811303619132, "resident_weight_gb": 23.509242304, "active_weight_gb": 4.201566656, "kv_alloc_gb_per_1k_tokens": 0.20992, "kv_read_gb_per_1k_tokens": 0.20992, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, model card metadata, gated base access check, linked GGUF file sizes, and a direct GGUF tensor-index range read. The authoritative bounds profile uses the API-selected Serenity-26B-A4B-HB16-Q6_K.gguf artifact: 23.509242304 GB resident, 2.915443136 GB fixed traffic, 0.160765440 GB per routed expert, 8 routed experts per token, 128 routed experts, 1 shared expert charged in fixed traffic, and layered Gemma 4 hybrid KV. Legacy per-1k KV fields show the short-context unsaturated coefficient; long-context calculation must use the profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/ReadyArt/Serenity-26B-A4B-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/ReadyArt/Serenity-26B-A4B-GGUF" }, { "label": "Base model card", "url": "https://huggingface.co/ReadyArt/Serenity-26B-A4B" }, { "label": "Selected HB16 Q6_K GGUF artifact", "url": "https://huggingface.co/ReadyArt/Serenity-26B-A4B-GGUF/resolve/644908e4cee7807801cc6363cfd5cc4c7240b19a/Serenity-26B-A4B-HB16-Q6_K.gguf" } ] }, { "id": "MaziyarPanahi/gemma-3-27b-it-GGUF", "name": "gemma-3-27b-it-GGUF Q2_K", "short_name": "Gemma 3 27B IT", "license": "gemma", "hf_pipeline_tag": "text-generation", "hf_downloads": 114543, "tags": [ "gguf", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:google/gemma-3-27b-it", "base_model:quantized:google/gemma-3-27b-it", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "gemma3 (GGUF Q2_K)", "total_params_b": 27.00900224, "active_params_b": 27.00900224, "layers": 62, "attention_heads": 32, "kv_heads": 16, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q2_K/Q3_K/Q6_K", "weight_bytes_per_param": 0.3888865131213377, "resident_weight_gb": 10.503436704, "active_weight_gb": 10.49687552, "kv_alloc_gb_per_1k_tokens": 0.08628207616, "kv_read_gb_per_1k_tokens": 0.095551488, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, model card metadata, package config absence check, gated base access check, linked GGUF file sizes, Google Gemma 3 local/global attention documentation, and a direct GGUF tensor-index range read. The HF API gguf.totalFileSize selects gemma-3-27b-it.Q2_K.gguf, so this row models the Q2_K artifact rather than larger GGUF siblings. Exact selected file residency is 10.503436704 GB and ordinary swept tensor traffic is 10.496875520 GB. KV uses Gemma 3 layered FP16 K/V over 10 full-context layers plus 52 sliding-window layers; the legacy catalog coefficients are effective display coefficients for the default compare contexts, while the profile contains the authoritative nonlinear layered KV adapter." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/gemma-3-27b-it-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/gemma-3-27b-it-GGUF" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-3-27b-it" }, { "label": "Selected Q2_K GGUF artifact", "url": "https://huggingface.co/MaziyarPanahi/gemma-3-27b-it-GGUF/resolve/f1a115f273da292b54e6ef203f5c4f8645f87d5f/gemma-3-27b-it.Q2_K.gguf" } ] }, { "id": "google/gemma-2b", "name": "gemma-2b BF16", "short_name": "Gemma 2B", "license": "gemma", "hf_pipeline_tag": "text-generation", "hf_downloads": 117744, "tags": [ "transformers", "safetensors", "gguf", "gemma", "text-generation", "arxiv:2312.11805", "arxiv:2009.03300", "arxiv:1905.07830", "arxiv:1911.11641", "arxiv:1904.09728", "arxiv:1905.10044", "arxiv:1907.10641", "arxiv:1811.00937", "arxiv:1809.02789", "arxiv:1911.01547", "arxiv:1705.03551", "arxiv:2107.03374", "arxiv:2108.07732", "arxiv:2110.14168", "arxiv:2304.06364", "arxiv:2206.04615", "arxiv:1804.06876", "arxiv:2110.08193", "arxiv:2009.11462", "arxiv:2101.11718", "arxiv:1804.09301", "arxiv:2109.07958", "arxiv:2203.09509", "license:gemma", "eval-results", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma (gated config unavailable)", "total_params_b": 2.506172416, "active_params_b": 2.506172416, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": 8192, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 5.012344832, "active_weight_gb": 5.012344832, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manually checked on 2026-07-07 from current Hugging Face API metadata, accessible model card text, and raw/HF CLI gated-access checks. The API exposes GemmaForCausalLM/model_type gemma and BF16 safetensors total 2506172416; the model card records an 8192-token context length; config, safetensors index/shards, and the GGUF object are gated from the configured osolmaz token, so KV geometry, tied embeddings, and swept decode traffic are not verified. The authoritative profile fails closed with unsupported-gated-config." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-2b" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-2b" }, { "label": "Pinned gated config access check", "url": "https://huggingface.co/google/gemma-2b/raw/9cf48e52b224239de00d483ec8eb84fb8d0f3a3a/config.json" } ] }, { "id": "MaziyarPanahi/Ministral-3-3B-Reasoning-2512-GGUF", "name": "Ministral-3-3B-Reasoning-2512-GGUF FP16", "short_name": "Ministral 3 3B Reasoning 2512", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 113921, "tags": [ "gguf", "mistral", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:mistralai/Ministral-3-3B-Reasoning-2512", "base_model:quantized:mistralai/Ministral-3-3B-Reasoning-2512", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "mistral3 (GGUF FP16)", "total_params_b": 3.429006336, "active_params_b": 3.429006336, "layers": 26, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF FP16/F32", "weight_bytes_per_param": 2, "resident_weight_gb": 6.8662192, "active_weight_gb": 6.858338304, "kv_alloc_gb_per_1k_tokens": 0.106496, "kv_read_gb_per_1k_tokens": 0.106496, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, model card, base Ministral3 config/API metadata, linked GGUF file sizes, and a direct selected FP16 GGUF tensor-index range read. The authoritative bounds profile uses Ministral-3-3B-Reasoning-2512.fp16.gguf because it matches API gguf.totalFileSize: 6.866219200 GB resident and 6.858338304 GB ordinary text-decode swept traffic. The selected text GGUF has tied embeddings and no output.weight tensor, so token_embd.weight is charged as output projection traffic. KV uses full-context FP16 K/V over 26 text layers." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/Ministral-3-3B-Reasoning-2512-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/MaziyarPanahi/Ministral-3-3B-Reasoning-2512-GGUF/raw/7c4a77bf3c700c12db2a3aab1e73fd87ed1d3b75/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/mistralai/Ministral-3-3B-Reasoning-2512" }, { "label": "Base config", "url": "https://huggingface.co/mistralai/Ministral-3-3B-Base-2512/raw/6f9c4b12a95b139af68670a6713616b757923735/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Ministral-3-3B-Reasoning-2512-GGUF" }, { "label": "Selected FP16 GGUF", "url": "https://huggingface.co/MaziyarPanahi/Ministral-3-3B-Reasoning-2512-GGUF/resolve/7c4a77bf3c700c12db2a3aab1e73fd87ed1d3b75/Ministral-3-3B-Reasoning-2512.fp16.gguf" } ] }, { "id": "MaziyarPanahi/solar-pro-preview-instruct-GGUF", "name": "solar-pro-preview-instruct-GGUF FP16", "short_name": "Solar Pro Preview Instruct", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 111877, "tags": [ "gguf", "mistral", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:upstage/solar-pro-preview-instruct", "base_model:quantized:upstage/solar-pro-preview-instruct", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "llama GGUF / SolarForCausalLM", "total_params_b": 22.140032, "active_params_b": 21.97553664, "layers": 64, "attention_heads": 40, "kv_heads": 10, "head_dim": 128, "max_context_tokens": 4096, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF FP16/F32", "weight_bytes_per_param": 2, "resident_weight_gb": 44.282148416, "active_weight_gb": 43.95239424, "kv_alloc_gb_per_1k_tokens": 0.32768, "kv_read_gb_per_1k_tokens": 0.32768, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, model card, base Solar config/API metadata, custom Solar code review, linked GGUF file sizes, and a direct selected FP16 GGUF tensor-index range read. The authoritative bounds profile uses solar-pro-preview-instruct.fp16.gguf because it matches API gguf.totalFileSize: 44.282148416 GB resident and 43.952394240 GB ordinary text-decode swept traffic after excluding token_embd.weight plus GGUF metadata/header/file overhead. The base config has sliding_window 2047, but the selected GGUF header lacks llama.attention.sliding_window, so this selected-artifact profile charges full-context FP16 K/V over 64 layers." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/solar-pro-preview-instruct-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/MaziyarPanahi/solar-pro-preview-instruct-GGUF/raw/4f05bde30de1ef0b38a2c3ad1e0f1480ef7156ff/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/upstage/solar-pro-preview-instruct" }, { "label": "Base config", "url": "https://huggingface.co/upstage/solar-pro-preview-instruct/raw/dd4bcf7006df9b1ce3f87711e702e4063832aae3/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/solar-pro-preview-instruct-GGUF" }, { "label": "Selected FP16 GGUF", "url": "https://huggingface.co/MaziyarPanahi/solar-pro-preview-instruct-GGUF/resolve/4f05bde30de1ef0b38a2c3ad1e0f1480ef7156ff/solar-pro-preview-instruct.fp16.gguf" } ] }, { "id": "MaziyarPanahi/mistral-small-3.1-24b-instruct-2503-hf-GGUF", "name": "mistral-small-3.1-24b-instruct-2503-hf-GGUF F16", "short_name": "mistral small 3.1 24b instruct 2503 hf", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 113869, "tags": [ "gguf", "mistral", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:mrfakename/mistral-small-3.1-24b-instruct-2503-hf", "base_model:quantized:mrfakename/mistral-small-3.1-24b-instruct-2503-hf", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "llama GGUF / MistralForCausalLM", "total_params_b": 23.5724, "active_params_b": 23.5724, "layers": 40, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF FP16/F32", "weight_bytes_per_param": 2, "resident_weight_gb": 47.153519136, "active_weight_gb": 45.80345856, "kv_alloc_gb_per_1k_tokens": 0.16384, "kv_read_gb_per_1k_tokens": 0.16384, "notes": "Audited on 2026-07-07 from current Hugging Face API metadata, base-wrapper config, selected linked-object sizes, and selected FP16 GGUF header. The API-selected artifact is mistral-small-3.1-24b-instruct-2503-hf.fp16.gguf; ordinary text decode excludes resident-only token_embd.weight and GGUF header/tokenizer overhead." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/mistral-small-3.1-24b-instruct-2503-hf-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/MaziyarPanahi/mistral-small-3.1-24b-instruct-2503-hf-GGUF/raw/69f1da83040e31989aa85ae485e3dff329d954ee/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/mrfakename/mistral-small-3.1-24b-instruct-2503-hf" }, { "label": "Base config", "url": "https://huggingface.co/mrfakename/mistral-small-3.1-24b-instruct-2503-hf/raw/822b1dd72c3f51176c7519290f1882593d5a10b0/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/mistral-small-3.1-24b-instruct-2503-hf-GGUF" }, { "label": "Selected FP16 GGUF", "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" } ] }, { "id": "google/gemma-4-31B-it-qat-q4_0-unquantized-assistant", "name": "Gemma 4 31B QAT Q4_0 Unquantized Assistant Drafter BF16", "short_name": "Gemma 4 31B QAT Assistant", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 143335, "tags": [ "transformers", "safetensors", "gemma4_assistant", "text-generation", "image-text-to-text", "conversational", "base_model:google/gemma-4-31B-it-assistant", "base_model:finetune:google/gemma-4-31B-it-assistant", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_assistant", "total_params_b": 0.469518596, "active_params_b": 0.469518596, "layers": 4, "attention_heads": 32, "kv_heads": 16, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 0.939037192, "active_weight_gb": 0.939037192, "kv_alloc_gb_per_1k_tokens": 0.053248, "kv_read_gb_per_1k_tokens": 0.053248, "notes": "Manually reviewed on 2026-07-07 from current Hugging Face API metadata, served config, model card, Transformers Gemma4Assistant source/docs, and direct safetensors header range read. This repo is a Gemma 4 MTP speculative drafter derived from google/gemma-4-31B-it-assistant. The profile records the exact 0.939037192 GB resident BF16 drafter artifact and intentionally disables production throughput because Bounds Engine v1 lacks paired target+assistant speculative decoding." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-4-31B-it-qat-q4_0-unquantized-assistant" }, { "label": "Model config", "url": "https://huggingface.co/google/gemma-4-31B-it-qat-q4_0-unquantized-assistant/raw/2194ebe30f476ea96d5e65503695dea6d78791bf/config.json" }, { "label": "Generation config", "url": "https://huggingface.co/google/gemma-4-31B-it-qat-q4_0-unquantized-assistant/raw/2194ebe30f476ea96d5e65503695dea6d78791bf/generation_config.json" }, { "label": "Base assistant model card", "url": "https://huggingface.co/google/gemma-4-31B-it-assistant" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-31B-it-qat-q4_0-unquantized-assistant" }, { "label": "Transformers Gemma4Assistant source", "url": "https://github.com/huggingface/transformers/blob/08a7ef05bcf9723cb2e58855afb8dc2c799323ff/src/transformers/models/gemma4_assistant/modeling_gemma4_assistant.py" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/google/gemma-4-31B-it-qat-q4_0-unquantized-assistant/resolve/2194ebe30f476ea96d5e65503695dea6d78791bf/model.safetensors" } ] }, { "id": "JunHowie/Qwen3-14B-GPTQ-Int4", "name": "Qwen3-14B-GPTQ-Int4 4-bit", "short_name": "Qwen3 14B GPTQ Int4", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 184023, "tags": [ "transformers", "safetensors", "qwen3", "text-generation", "Qwen3", "GPTQ", "Int4", "\u91cf\u5316\u4fee\u590d", "vLLM", "conversational", "arxiv:2309.00071", "arxiv:2505.09388", "base_model:Qwen/Qwen3-14B", "base_model:quantized:Qwen/Qwen3-14B", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "4-bit", "gptq", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 14.7683, "active_params_b": 14.7683, "layers": 40, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GPTQ Int4/BF16/F16", "weight_bytes_per_param": 0.6760609963476383, "resident_weight_gb": 9.98427648, "active_weight_gb": 8.42845184, "kv_alloc_gb_per_1k_tokens": 0.16384, "kv_read_gb_per_1k_tokens": 0.16384, "notes": "Audited on 2026-07-07 from current Hugging Face API metadata, served GPTQ config, quantize_config, base Qwen3-14B config comparison, safetensors index, linked-object sizes, and direct shard header range reads. Ordinary text decode excludes resident-only model.embed_tokens.weight and includes quantized layer tensors, norms, and unquantized lm_head.weight." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/JunHowie/Qwen3-14B-GPTQ-Int4" }, { "label": "Model config", "url": "https://huggingface.co/JunHowie/Qwen3-14B-GPTQ-Int4/raw/5073a3e15216fb4facc39c953934910dbf0df51b/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-14B" }, { "label": "Base config", "url": "https://huggingface.co/Qwen/Qwen3-14B/raw/40c069824f4251a91eefaf281ebe4c544efd3e18/config.json" }, { "label": "Quantize config", "url": "https://huggingface.co/JunHowie/Qwen3-14B-GPTQ-Int4/raw/5073a3e15216fb4facc39c953934910dbf0df51b/quantize_config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/JunHowie/Qwen3-14B-GPTQ-Int4" }, { "label": "Safetensors index", "url": "https://huggingface.co/JunHowie/Qwen3-14B-GPTQ-Int4/raw/5073a3e15216fb4facc39c953934910dbf0df51b/model.safetensors.index.json" } ] }, { "id": "MaziyarPanahi/Llama-3-8B-Instruct-64k-GGUF", "name": "Llama-3-8B-Instruct-64k-GGUF FP16", "short_name": "Llama 3 8B Instruct 64k", "license": "llama3", "hf_pipeline_tag": "text-generation", "hf_downloads": 113638, "tags": [ "transformers", "gguf", "mistral", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "llama", "llama-3", "base_model:MaziyarPanahi/Llama-3-8B-Instruct-64k", "base_model:quantized:MaziyarPanahi/Llama-3-8B-Instruct-64k", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "llama GGUF / LlamaForCausalLM", "total_params_b": 8.030261248, "active_params_b": 8.030261248, "layers": 32, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 64000, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF FP16/F32", "weight_bytes_per_param": 2, "resident_weight_gb": 16.068890848, "active_weight_gb": 15.010381824, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "notes": "Audited on 2026-07-07 from current Hugging Face API metadata, base model card/config, selected linked-object sizes, and selected FP16 GGUF header. The API-selected artifact is Llama-3-8B-Instruct-64k.fp16.gguf; ordinary text decode excludes resident-only token_embd.weight and GGUF header/tokenizer overhead." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/Llama-3-8B-Instruct-64k-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/MaziyarPanahi/Llama-3-8B-Instruct-64k-GGUF/raw/3785b79d1c23332e60859b853ed9807f94844353/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/MaziyarPanahi/Llama-3-8B-Instruct-64k" }, { "label": "Base config", "url": "https://huggingface.co/MaziyarPanahi/Llama-3-8B-Instruct-64k/raw/7423c85617da4b946d251274c6958d9533fbbc5c/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Llama-3-8B-Instruct-64k-GGUF" }, { "label": "Selected FP16 GGUF", "url": "https://huggingface.co/MaziyarPanahi/Llama-3-8B-Instruct-64k-GGUF/resolve/3785b79d1c23332e60859b853ed9807f94844353/Llama-3-8B-Instruct-64k.fp16.gguf" } ] }, { "id": "zai-org/GLM-5.2", "name": "GLM-5.2 BF16/F32", "short_name": "GLM 5.2", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 231218, "tags": [ "transformers", "safetensors", "glm_moe_dsa", "text-generation", "conversational", "en", "zh", "arxiv:2602.15763", "arxiv:2603.12201", "license:mit", "eval-results", "endpoints_compatible", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "glm_moe_dsa (GlmMoeDsaForCausalLM)", "total_params_b": 753.32994048, "active_params_b": 23.5416, "layers": 78, "attention_heads": 64, "kv_heads": 64, "head_dim": 256, "max_context_tokens": 1048576, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "BF16/F32", "weight_bytes_per_param": 2.00000005165333, "resident_weight_gb": 1506.659919872, "active_weight_gb": 61.874560512, "kv_alloc_gb_per_1k_tokens": 5.117184, "kv_read_gb_per_1k_tokens": 5.117184, "expert_param_b": 2.868903861905781, "fixed_param_b": 7.9860485617469985, "shared_expert_param_b": 0, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, pinned config comparison with GLM-5.2-FP8, direct safetensors header range reads across all 282 shards, and the Transformers GLM MoE DSA config/runtime/cache path. Active traffic is the single-session expected-distinct MoE text-decode weight traffic. KV display fields charge expanded BF16 K/V for 78 layers plus GLM-5.2 IndexShare DSA indexer key cache for 21 full indexer layers. The audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/zai-org/GLM-5.2" }, { "label": "Model config", "url": "https://huggingface.co/zai-org/GLM-5.2/raw/b4734de4facf877f85769a911abafc5283eab3d9/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/zai-org/GLM-5.2" }, { "label": "Safetensors index", "url": "https://huggingface.co/zai-org/GLM-5.2/resolve/b4734de4facf877f85769a911abafc5283eab3d9/model.safetensors.index.json" } ] }, { "id": "MaziyarPanahi/Qwen2-7B-Instruct-GGUF", "name": "Qwen2-7B-Instruct-GGUF GGUF FP16", "short_name": "Qwen2 7B Instruct", "license": "unknown", "hf_pipeline_tag": "text-generation", "hf_downloads": 113445, "tags": [ "transformers", "gguf", "mistral", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "llama-3", "llama", "base_model:Qwen/Qwen2-7B-Instruct", "base_model:quantized:Qwen/Qwen2-7B-Instruct", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 GGUF (Qwen2ForCausalLM)", "total_params_b": 7.615616512, "active_params_b": 7.070619136, "layers": 28, "attention_heads": 28, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF FP16", "weight_bytes_per_param": 2, "resident_weight_gb": 15.237850656, "active_weight_gb": 14.141904896, "kv_alloc_gb_per_1k_tokens": 0.057344, "kv_read_gb_per_1k_tokens": 0.057344, "notes": "Audited on 2026-07-07 from current Hugging Face API metadata, base model API/config metadata, selected linked-object sizes, and selected FP16 GGUF header. The package config is a stale Mistral stub and is not used as architecture evidence. The API-selected artifact is Qwen2-7B-Instruct.fp16.gguf; ordinary text decode excludes resident-only token_embd.weight and GGUF header/tokenizer overhead." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/Qwen2-7B-Instruct-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/MaziyarPanahi/Qwen2-7B-Instruct-GGUF/raw/8cd289397d4eb8a1faa022adb9bbab5aae2d1739/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2-7B-Instruct" }, { "label": "Base config", "url": "https://huggingface.co/Qwen/Qwen2-7B-Instruct/raw/f2826a00ceef68f0f2b946d945ecc0477ce4450c/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Qwen2-7B-Instruct-GGUF" }, { "label": "Selected FP16 GGUF", "url": "https://huggingface.co/MaziyarPanahi/Qwen2-7B-Instruct-GGUF/resolve/8cd289397d4eb8a1faa022adb9bbab5aae2d1739/Qwen2-7B-Instruct.fp16.gguf" } ] }, { "id": "MaziyarPanahi/Yi-1.5-6B-Chat-GGUF", "name": "Yi-1.5-6B-Chat-GGUF GGUF Q2_K", "short_name": "Yi 1.5 6B Chat", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 112573, "tags": [ "transformers", "gguf", "mistral", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "safetensors", "llama", "text-generation", "conversational", "arxiv:2403.04652", "license:apache-2.0", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "region:us", "base_model:01-ai/Yi-1.5-6B-Chat", "base_model:quantized:01-ai/Yi-1.5-6B-Chat" ], "architecture": { "type": "Dense transformer", "detail": "llama GGUF (LlamaForCausalLM)", "total_params_b": 6.06103552, "active_params_b": 5.79889152, "layers": 32, "attention_heads": 32, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 4096, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q2_K", "weight_bytes_per_param": 0.38558871207539136, "resident_weight_gb": 2.33706688, "active_weight_gb": 2.249555968, "kv_alloc_gb_per_1k_tokens": 0.065536, "kv_read_gb_per_1k_tokens": 0.065536, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, model card, pinned base config comparison, linked GGUF file sizes, stale config check, and direct selected Q2_K GGUF tensor-index range read. The authoritative bounds profile uses Yi-1.5-6B-Chat.Q2_K.gguf: 2.337066880 GB resident and 2.249555968 GB ordinary text-decode swept traffic after excluding token_embd.weight." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/Yi-1.5-6B-Chat-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/MaziyarPanahi/Yi-1.5-6B-Chat-GGUF/raw/5eeb4c393e2a235fd5dce7f6007dd5ea96577592/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Yi-1.5-6B-Chat-GGUF" }, { "label": "Selected GGUF tensor-index audit", "url": "https://huggingface.co/MaziyarPanahi/Yi-1.5-6B-Chat-GGUF/resolve/5eeb4c393e2a235fd5dce7f6007dd5ea96577592/Yi-1.5-6B-Chat.Q2_K.gguf" }, { "label": "Base model card", "url": "https://huggingface.co/01-ai/Yi-1.5-6B-Chat" }, { "label": "Base model config", "url": "https://huggingface.co/01-ai/Yi-1.5-6B-Chat/raw/771924d1c83d67527d665913415d7086f11ea9c0/config.json" } ] }, { "id": "MaziyarPanahi/mathstral-7B-v0.1-GGUF", "name": "mathstral-7B-v0.1-GGUF GGUF FP16", "short_name": "mathstral 7B v0.1", "license": "unknown", "hf_pipeline_tag": "text-generation", "hf_downloads": 113042, "tags": [ "transformers", "gguf", "mistral", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "region:us", "base_model:mistralai/Mathstral-7B-v0.1", "base_model:quantized:mistralai/Mathstral-7B-v0.1" ], "architecture": { "type": "Dense transformer", "detail": "llama GGUF / MistralForCausalLM", "total_params_b": 7.248023552, "active_params_b": 7.113805824, "layers": 32, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF F16/F32", "weight_bytes_per_param": 2, "resident_weight_gb": 14.497336928, "active_weight_gb": 14.228144128, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, model card, base Mathstral config/API metadata, linked GGUF file sizes, stale package config check, and direct selected F16 GGUF tensor-index range read. The authoritative bounds profile uses mathstral-7B-v0.1.fp16.gguf because it matches API gguf.totalFileSize: 14.497336928 GB resident and 14.228144128 GB ordinary text-decode swept traffic after excluding token_embd.weight plus GGUF metadata/header/file overhead." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/mathstral-7B-v0.1-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/MaziyarPanahi/mathstral-7B-v0.1-GGUF/raw/d3c8fafa377cf389d822360b1a289226ba904d7d/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/mathstral-7B-v0.1-GGUF" }, { "label": "Selected GGUF tensor-index audit", "url": "https://huggingface.co/MaziyarPanahi/mathstral-7B-v0.1-GGUF/resolve/d3c8fafa377cf389d822360b1a289226ba904d7d/mathstral-7B-v0.1.fp16.gguf" }, { "label": "Base model card", "url": "https://huggingface.co/mistralai/Mathstral-7B-v0.1" }, { "label": "Base model config", "url": "https://huggingface.co/mistralai/Mathstral-7B-v0.1/raw/ec3a48484ef241dfe03282edcb0f25e564923823/config.json" } ] }, { "id": "MaziyarPanahi/Mistral-Small-Instruct-2409-GGUF", "name": "Mistral-Small-Instruct-2409-GGUF GGUF FP16", "short_name": "Mistral Small Instruct 2409", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 113311, "tags": [ "gguf", "mistral", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:mistralai/Mistral-Small-Instruct-2409", "base_model:quantized:mistralai/Mistral-Small-Instruct-2409", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "llama GGUF / MistralForCausalLM", "total_params_b": 22.247282688, "active_params_b": 22.045956096, "layers": 56, "attention_heads": 48, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF F16/F32", "weight_bytes_per_param": 2, "resident_weight_gb": 44.496728832, "active_weight_gb": 44.093300736, "kv_alloc_gb_per_1k_tokens": 0.229376, "kv_read_gb_per_1k_tokens": 0.229376, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, model card, base Mistral Small config/API metadata, linked GGUF file sizes, stale package config check, and direct selected F16 GGUF tensor-index range read. The authoritative bounds profile uses Mistral-Small-Instruct-2409.fp16.gguf because it matches API gguf.totalFileSize: 44.496728832 GB resident and 44.093300736 GB ordinary text-decode swept traffic after excluding token_embd.weight plus GGUF metadata/header/file overhead." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/Mistral-Small-Instruct-2409-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/MaziyarPanahi/Mistral-Small-Instruct-2409-GGUF/raw/23bcb4eba278cea53f211652d60dbcd6724f89e2/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Mistral-Small-Instruct-2409-GGUF" }, { "label": "Selected GGUF tensor-index audit", "url": "https://huggingface.co/MaziyarPanahi/Mistral-Small-Instruct-2409-GGUF/resolve/23bcb4eba278cea53f211652d60dbcd6724f89e2/Mistral-Small-Instruct-2409.fp16.gguf" }, { "label": "Base model card", "url": "https://huggingface.co/mistralai/Mistral-Small-Instruct-2409" }, { "label": "Base model config", "url": "https://huggingface.co/mistralai/Mistral-Small-Instruct-2409/raw/4600506f6b13c7ef89e61a54263f4c9bf483de30/config.json" } ] }, { "id": "moonshotai/Kimi-VL-A3B-Thinking", "name": "Kimi-VL-A3B-Thinking BF16", "short_name": "Kimi VL A3B Thinking", "license": "mit", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 145860, "tags": [ "transformers", "safetensors", "kimi_vl", "feature-extraction", "image-text-to-text", "conversational", "custom_code", "arxiv:2504.07491", "base_model:moonshotai/Kimi-VL-A3B-Instruct", "base_model:finetune:moonshotai/Kimi-VL-A3B-Instruct", "license:mit", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "kimi_vl with DeepseekV3 language model", "total_params_b": 16.407657776, "active_params_b": 2.5792336, "layers": 27, "attention_heads": 16, "kv_heads": 16, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": 64, "routed_experts_per_token": 6, "shared_experts_per_token": 2 }, "adapter": { "kind": "moe", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 32.815315552, "active_weight_gb": 5.1584672, "kv_alloc_gb_per_1k_tokens": 0.27648, "kv_read_gb_per_1k_tokens": 0.27648, "expert_param_b": 0.224919552, "fixed_param_b": 1.229716288, "shared_expert_param_b": 0, "notes": "Manual correction on 2026-07-07 from current Hugging Face API metadata, pinned served config, Kimi-VL-A3B-Instruct config/runtime comparison, model card, custom runtime diff review, safetensors index, and direct range-read safetensors shard headers. Legacy fallback values mirror the audited profile: exact BF16 resident bytes, resident-only vision/projector/input-embedding tensors, fixed always-on language/shared-expert traffic, expected distinct routed experts, and expanded BF16 K/V cache from the custom DeepSeekV3 attention path." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/moonshotai/Kimi-VL-A3B-Thinking" }, { "label": "Model config", "url": "https://huggingface.co/moonshotai/Kimi-VL-A3B-Thinking/raw/7d99e220af610d8624fcba22b2c076c7ed528f14/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/moonshotai/Kimi-VL-A3B-Thinking" }, { "label": "Safetensors index audit", "url": "https://huggingface.co/moonshotai/Kimi-VL-A3B-Thinking/resolve/7d99e220af610d8624fcba22b2c076c7ed528f14/model.safetensors.index.json" }, { "label": "Base model card", "url": "https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct" } ] }, { "id": "MaziyarPanahi/WizardLM-2-7B-GGUF", "name": "WizardLM-2-7B-GGUF GGUF FP16", "short_name": "WizardLM 2 7B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 112698, "tags": [ "transformers", "gguf", "mistral", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "safetensors", "text-generation", "arxiv:2304.12244", "arxiv:2306.08568", "arxiv:2308.09583", "license:apache-2.0", "autotrain_compatible", "endpoints_compatible", "text-generation-inference", "region:us", "base_model:microsoft/WizardLM-2-7B", "base_model:quantized:microsoft/WizardLM-2-7B" ], "architecture": { "type": "Dense transformer", "detail": "llama GGUF / Mistral-derived", "total_params_b": 7.241732096, "active_params_b": 7.110660096, "layers": 32, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF F16/F32", "weight_bytes_per_param": 2, "resident_weight_gb": 14.484731552, "active_weight_gb": 14.221852672, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, model card, inaccessible base config check, stale package config check, linked GGUF file sizes, and direct selected F16 GGUF tensor-index range read. The authoritative bounds profile uses WizardLM-2-7B.fp16.gguf because it matches API gguf.totalFileSize: 14.484731552 GB resident and 14.221852672 GB ordinary text-decode swept traffic after excluding token_embd.weight plus GGUF metadata/header/file overhead." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/WizardLM-2-7B-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/MaziyarPanahi/WizardLM-2-7B-GGUF/raw/491e4644eb4874621cb7d00153a20e0bf7358fb8/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/WizardLM-2-7B-GGUF" }, { "label": "Selected GGUF tensor-index audit", "url": "https://huggingface.co/MaziyarPanahi/WizardLM-2-7B-GGUF/resolve/491e4644eb4874621cb7d00153a20e0bf7358fb8/WizardLM-2-7B.fp16.gguf" }, { "label": "Base config access check", "url": "https://huggingface.co/microsoft/WizardLM-2-7B/raw/main/config.json" } ] }, { "id": "nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16", "name": "NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 BF16/F32", "short_name": "NVIDIA Nemotron 3 Ultra 550B A55B BF16", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 110467, "tags": [ "transformers", "safetensors", "nemotron_h", "text-generation", "nvidia", "pytorch", "nemotron-3", "latent-moe", "mtp", "conversational", "en", "fr", "es", "it", "de", "pt", "ja", "ko", "hi", "ar", "zh", "he", "dataset:nvidia/nemotron-post-training-v3", "dataset:nvidia/nemotron-pre-training-datasets", "license:other", "eval-results", "endpoints_compatible", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "nemotron_h (NemotronHForCausalLM)", "total_params_b": 550, "active_params_b": 55, "layers": 108, "attention_heads": 64, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 512, "routed_experts_per_token": 22, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "BF16/F32", "weight_bytes_per_param": 2, "resident_weight_gb": 1121.049257984, "active_weight_gb": 109.970251776, "kv_alloc_gb_per_1k_tokens": 0.006144, "kv_read_gb_per_1k_tokens": 0.006144, "kv_alloc_fixed_gb": 0.20840448, "kv_read_fixed_gb": 0.20840448, "expert_param_b": 1.00663296, "fixed_param_b": 32.839200768, "shared_expert_param_b": 0, "notes": "Manual correction on 2026-07-07 from current Hugging Face API metadata, pinned model card, pinned served config, BF16 base config comparison, Nemotron-H runtime cache review, safetensors index, and direct range-read safetensors shard headers. Legacy fallback values mirror the audited profile: exact stored BF16/F32 resident bytes, fixed always-on hybrid/shared-expert traffic, expected distinct routed experts, FP8 attention KV, and fixed float16-sized Mamba recurrent state. The authoritative profile supersedes the generated 1121.0492 GB resident estimate and zero-KV placeholders." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16" }, { "label": "Pinned served config", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16/raw/624ba927cfbef0427354998700de3d51173c8c04/config.json" }, { "label": "Pinned BF16 base config comparison", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-Base-BF16/raw/1493b9c9ca301d6b6e5b0bee90fef1990ba49ed4/config.json" }, { "label": "Nemotron-H runtime cache reference", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4/raw/4f0cf9daaeb7a4d5e23f80a00e7ed15f0e03caf6/modeling_nemotron_h.py" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16/resolve/624ba927cfbef0427354998700de3d51173c8c04/model.safetensors.index.json" } ] }, { "id": "MaziyarPanahi/QwQ-32B-GGUF", "name": "QwQ-32B-GGUF GGUF Q2_K", "short_name": "QwQ 32B Q2_K", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 112246, "tags": [ "gguf", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:Qwen/QwQ-32B", "base_model:quantized:Qwen/QwQ-32B", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "llama GGUF / Qwen2-derived", "total_params_b": 32.763876352, "active_params_b": 32.763876352, "layers": 64, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q2_K", "weight_bytes_per_param": 0.37583182831320683, "resident_weight_gb": 12.313707552, "active_weight_gb": 12.051652608, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Manual correction on 2026-07-07 from current Hugging Face API metadata, model card, pinned Qwen base API/config metadata, package config absence check, linked GGUF file sizes, and direct selected Q2_K GGUF tensor-index range read. The authoritative bounds profile uses QwQ-32B.Q2_K.gguf because it matches API gguf.totalFileSize: 12.313707552 GB resident and 12.051652608 GB ordinary text-decode swept traffic after excluding token_embd.weight plus GGUF metadata/header/file overhead." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/QwQ-32B-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/QwQ-32B-GGUF" }, { "label": "Base model API metadata", "url": "https://huggingface.co/api/models/Qwen/QwQ-32B" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/QwQ-32B/raw/976055f8c83f394f35dbd3ab09a285a984907bd0/config.json" }, { "label": "Selected GGUF tensor-index audit", "url": "https://huggingface.co/MaziyarPanahi/QwQ-32B-GGUF/resolve/c289541240c414b14fc9a8b2d48fac949dc16012/QwQ-32B.Q2_K.gguf" }, { "label": "Package config absence check", "url": "https://huggingface.co/MaziyarPanahi/QwQ-32B-GGUF/raw/c289541240c414b14fc9a8b2d48fac949dc16012/config.json" } ] }, { "id": "MaziyarPanahi/phi-4-GGUF", "name": "phi-4-GGUF GGUF FP16", "short_name": "phi 4 FP16", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 112039, "tags": [ "gguf", "mistral", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:microsoft/phi-4", "base_model:quantized:microsoft/phi-4", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "phi3 GGUF (Phi3ForCausalLM)", "total_params_b": 14.6595072, "active_params_b": 14.14570496, "layers": 40, "attention_heads": 40, "kv_heads": 10, "head_dim": 128, "max_context_tokens": 16384, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF FP16", "weight_bytes_per_param": 2, "resident_weight_gb": 29.323399264, "active_weight_gb": 28.29223936, "kv_alloc_gb_per_1k_tokens": 0.2048, "kv_read_gb_per_1k_tokens": 0.2048, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, model card, audited BF16 base profile comparison, linked GGUF file sizes, stale config check, and direct selected FP16 GGUF tensor-index range read. The authoritative bounds profile uses phi-4.fp16.gguf because it matches API gguf.totalFileSize: 29.323399264 GB resident and 28.292239360 GB ordinary text-decode swept traffic after excluding token_embd.weight plus GGUF metadata/header/file overhead." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/phi-4-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/MaziyarPanahi/phi-4-GGUF/raw/7b3c692e11c2028dd77bd2c22ecc651fb1a32261/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/phi-4-GGUF" }, { "label": "Base model card", "url": "https://huggingface.co/microsoft/phi-4" }, { "label": "Selected GGUF tensor-index audit", "url": "https://huggingface.co/MaziyarPanahi/phi-4-GGUF/resolve/7b3c692e11c2028dd77bd2c22ecc651fb1a32261/phi-4.fp16.gguf" } ] }, { "id": "LGAI-EXAONE/EXAONE-3.5-32B-Instruct-AWQ", "name": "EXAONE-3.5-32B-Instruct-AWQ 4-bit", "short_name": "EXAONE 3.5 32B Instruct", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 157661, "tags": [ "transformers", "safetensors", "exaone", "text-generation", "lg-ai", "exaone-3.5", "conversational", "custom_code", "en", "ko", "arxiv:2412.04862", "license:other", "4-bit", "awq", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "exaone (ExaoneForCausalLM)", "total_params_b": 32.0032, "active_params_b": 31.478912, "layers": 64, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "4-bit", "weight_bytes_per_param": 0.5, "resident_weight_gb": 18.18002432, "active_weight_gb": 17.13144832, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Manual correction on 2026-07-07 from current Hugging Face API metadata, model card, served AWQ config, base config comparison, custom runtime cache review, license file, safetensors index, and direct range-read shard headers. Bound weights use exact stored AWQ bytes: 18.180024320 GB resident and 17.131448320 GB ordinary text-decode swept traffic after excluding transformer.wte.weight." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/LGAI-EXAONE/EXAONE-3.5-32B-Instruct-AWQ" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/LGAI-EXAONE/EXAONE-3.5-32B-Instruct-AWQ" }, { "label": "Model config", "url": "https://huggingface.co/LGAI-EXAONE/EXAONE-3.5-32B-Instruct-AWQ/raw/278e45157895e03158845c2c873ef6a37bd15a0f/config.json" }, { "label": "Base config comparison", "url": "https://huggingface.co/LGAI-EXAONE/EXAONE-3.5-32B-Instruct/raw/3d27aab477c042df2e174d36eea47b77c8871ef0/config.json" }, { "label": "Custom runtime cache review", "url": "https://huggingface.co/LGAI-EXAONE/EXAONE-3.5-32B-Instruct-AWQ/raw/278e45157895e03158845c2c873ef6a37bd15a0f/modeling_exaone.py" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/LGAI-EXAONE/EXAONE-3.5-32B-Instruct-AWQ/resolve/278e45157895e03158845c2c873ef6a37bd15a0f/model.safetensors.index.json" } ] }, { "id": "microsoft/Phi-3.5-MoE-instruct", "name": "Phi-3.5-MoE-instruct BF16", "short_name": "Phi 3.5 MoE", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 162753, "tags": [ "transformers", "safetensors", "phimoe", "text-generation", "nlp", "code", "conversational", "custom_code", "multilingual", "arxiv:2404.14219", "arxiv:2407.13833", "arxiv:2403.06412", "license:mit", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "phimoe (PhiMoEForCausalLM)", "total_params_b": 41.873153344, "active_params_b": 6.5096656, "layers": 32, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": 16, "routed_experts_per_token": 2, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 83.746306688, "active_weight_gb": 13.0193312, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "notes": "Manual correction on 2026-07-07 from current Hugging Face API metadata, model card, config, custom modeling code, license file, safetensors index, and direct range-read shard headers. Bound weights use exact BF16 tensor bytes: 83.746306688 GB resident, 2.953001600 GB fixed ordinary text-decode traffic, and 5.033164800 GB per routed expert index. Active traffic is fixed plus two routed experts: 13.019331200 GB." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/microsoft/Phi-3.5-MoE-instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/microsoft/Phi-3.5-MoE-instruct" }, { "label": "Model config", "url": "https://huggingface.co/microsoft/Phi-3.5-MoE-instruct/raw/43688451b462a3351d8580625ebe1931adb3986d/config.json" }, { "label": "Custom modeling code", "url": "https://huggingface.co/microsoft/Phi-3.5-MoE-instruct/raw/43688451b462a3351d8580625ebe1931adb3986d/modeling_phimoe.py" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/microsoft/Phi-3.5-MoE-instruct/resolve/43688451b462a3351d8580625ebe1931adb3986d/model.safetensors.index.json" } ] }, { "id": "MaziyarPanahi/Mistral-Large-Instruct-2411-GGUF", "name": "Mistral-Large-Instruct-2411-GGUF GGUF 2-bit", "short_name": "Mistral Large Instruct 2411", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 111457, "tags": [ "gguf", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:mistralai/Mistral-Large-Instruct-2411", "base_model:quantized:mistralai/Mistral-Large-Instruct-2411", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "llama GGUF / Mistral-derived", "total_params_b": 122.610069504, "active_params_b": 122.20741632, "layers": 88, "attention_heads": 96, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q2_K", "weight_bytes_per_param": 0.3686181604401222, "resident_weight_gb": 45.196298272, "active_weight_gb": 45.063389184, "kv_alloc_gb_per_1k_tokens": 0.360448, "kv_read_gb_per_1k_tokens": 0.360448, "notes": "Manual correction on 2026-07-07 from current Hugging Face API metadata, model card, pinned base config/API metadata, selected linked-object metadata, package config absence check, and direct selected Q2_K GGUF tensor-index range read. The authoritative bounds profile uses Mistral-Large-Instruct-2411.Q2_K.gguf because it matches API gguf.totalFileSize: 45.196298272 GB resident and 45.063389184 GB ordinary text-decode swept traffic after excluding token_embd.weight plus GGUF metadata/header/file overhead." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/Mistral-Large-Instruct-2411-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Mistral-Large-Instruct-2411-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/mistralai/Mistral-Large-Instruct-2411/raw/ba78820945ae22361b0274cf0ae6d696c967c1a4/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/mistralai/Mistral-Large-Instruct-2411" }, { "label": "Selected GGUF tensor-index audit", "url": "https://huggingface.co/MaziyarPanahi/Mistral-Large-Instruct-2411-GGUF/resolve/76081fc502213f8e894d6089bc648fa4fcc962c9/Mistral-Large-Instruct-2411.Q2_K.gguf" } ] }, { "id": "MaziyarPanahi/firefunction-v2-GGUF", "name": "firefunction-v2-GGUF GGUF 2-bit", "short_name": "firefunction v2", "license": "llama3", "hf_pipeline_tag": "text-generation", "hf_downloads": 111313, "tags": [ "transformers", "gguf", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "safetensors", "text-generation", "conversational", "function-calling", "text-generation-inference", "region:us", "base_model:fireworks-ai/llama-3-firefunction-v2", "base_model:quantized:fireworks-ai/llama-3-firefunction-v2", "license:llama3", "imatrix" ], "architecture": { "type": "Dense transformer", "detail": "llama (LlamaForCausalLM)", "total_params_b": 70.553706496, "active_params_b": 69.503033344, "layers": 80, "attention_heads": 64, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 8192, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF IQ1_M", "weight_bytes_per_param": 0.23742478857506516, "resident_weight_gb": 16.751198848, "active_weight_gb": 16.39858176, "kv_alloc_gb_per_1k_tokens": 0.32768, "kv_read_gb_per_1k_tokens": 0.32768, "notes": "Manual correction on 2026-07-07 from current Hugging Face API metadata, model card, pinned base config/API metadata, linked-object metadata, package config absence check, and direct selected IQ1_M GGUF tensor-index range read. The authoritative bounds profile uses firefunction-v2.IQ1_M.gguf because it matches API gguf.totalFileSize: 16.751198848 GB resident and 16.398581760 GB ordinary text-decode swept traffic after excluding token_embd.weight plus GGUF metadata/header/file overhead." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/firefunction-v2-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/firefunction-v2-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/fireworks-ai/llama-3-firefunction-v2/raw/7954e120254ced10ac26a2dd346679d7328fa9ae/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/fireworks-ai/llama-3-firefunction-v2" }, { "label": "Selected GGUF tensor-index audit", "url": "https://huggingface.co/MaziyarPanahi/firefunction-v2-GGUF/resolve/e48d670a217a6e6fbbc3d929dbd8cb0ea25d80f8/firefunction-v2.IQ1_M.gguf" } ] }, { "id": "microsoft/biogpt", "name": "BioGPT FP32", "short_name": "BioGPT", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 122777, "tags": [ "transformers", "pytorch", "biogpt", "text-generation", "en", "license:mit", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "biogpt (BioGptForCausalLM)", "total_params_b": 0.346763264, "active_params_b": 0.34571264, "layers": 24, "attention_heads": 16, "kv_heads": 16, "head_dim": 64, "max_context_tokens": 1024, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "FP32", "weight_bytes_per_param": 4, "resident_weight_gb": 1.387053056, "active_weight_gb": 1.38285056, "kv_alloc_gb_per_1k_tokens": 0.196608, "kv_read_gb_per_1k_tokens": 0.196608, "notes": "Manual correction on 2026-07-07 from current Hugging Face API metadata, pinned config and model card, Transformers BioGPT implementation review, PyTorch linked-object metadata, and range-read PyTorch zip data.pkl tensor metadata. The checkpoint stores 389 FloatStorage tensors; unique tied serving resident parameters total 1.387053056 GB, ordinary text decode sweeps 1.382850560 GB, and biogpt.embed_positions.weight is the 0.004202496 GB resident-only learned position table." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/microsoft/biogpt" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/microsoft/biogpt" }, { "label": "Pinned model config", "url": "https://huggingface.co/microsoft/biogpt/raw/eb0d815e95434dc9e3b78f464e52b899bee7d923/config.json" }, { "label": "Pinned PyTorch checkpoint metadata audit", "url": "https://huggingface.co/microsoft/biogpt/resolve/eb0d815e95434dc9e3b78f464e52b899bee7d923/pytorch_model.bin" } ] }, { "id": "MaziyarPanahi/INTELLECT-2-GGUF", "name": "INTELLECT-2-GGUF GGUF Q2_K", "short_name": "INTELLECT 2", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 110740, "tags": [ "gguf", "quantized", "2-bit", "3-bit", "4-bit", "5-bit", "6-bit", "8-bit", "GGUF", "text-generation", "base_model:PrimeIntellect/INTELLECT-2", "base_model:quantized:PrimeIntellect/INTELLECT-2", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 32.763876352, "active_params_b": 31.985308672, "layers": 64, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q2_K", "weight_bytes_per_param": 0.3758132527333987, "resident_weight_gb": 12.313098944, "active_weight_gb": 12.051652608, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Manual correction on 2026-07-07 from current Hugging Face API metadata, pinned base config/API metadata, package config absence check, linked-object HEAD checks, and direct selected Q2_K GGUF tensor-index range read. The API-selected INTELLECT-2.Q2_K.gguf linked file is 12.313098944 GB. Tensor payloads total 12.307120128 GB, ordinary text decode sweeps 12.051652608 GB after excluding token_embd.weight and GGUF metadata/header overhead, and full-context FP16 GGUF KV remains 0.262144 GB per 1k tokens." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/INTELLECT-2-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/INTELLECT-2-GGUF" }, { "label": "Pinned base model config", "url": "https://huggingface.co/PrimeIntellect/INTELLECT-2/raw/93e0fc8da282b7273d8d432878b626843d97d667/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/PrimeIntellect/INTELLECT-2" }, { "label": "Selected Q2_K GGUF tensor-index audit", "url": "https://huggingface.co/MaziyarPanahi/INTELLECT-2-GGUF/resolve/276790213c9cb201745d2e85061e6e374b2ee9f7/INTELLECT-2.Q2_K.gguf" } ] }, { "id": "Qwen/Qwen2.5-3B-Instruct-AWQ", "name": "Qwen2.5-3B-Instruct-AWQ 4-bit", "short_name": "Qwen2.5 3B Instruct", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 127257, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "chat", "conversational", "en", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-3B-Instruct", "base_model:quantized:Qwen/Qwen2.5-3B-Instruct", "license:other", "text-generation-inference", "endpoints_compatible", "4-bit", "awq", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM, AWQ GEMM)", "total_params_b": 3.397103616, "active_params_b": 3.085938688, "layers": 36, "attention_heads": 16, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "AWQ 4-bit", "weight_bytes_per_param": 0.7908499332626774, "resident_weight_gb": 2.686599168, "active_weight_gb": 2.064269312, "kv_alloc_gb_per_1k_tokens": 0.036864, "kv_read_gb_per_1k_tokens": 0.036864, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, pinned served AWQ config, pinned BF16 base config/API comparison, linked-object HEAD metadata, and a direct safetensors header range read. Bound weights separate the exact 2.686599168 GB resident AWQ payload from the 2.064269312 GB ordinary text-decode swept subset; the audited profile supersedes the generated ideal 4-bit estimate." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-3B-Instruct-AWQ" }, { "label": "Pinned AWQ model config", "url": "https://huggingface.co/Qwen/Qwen2.5-3B-Instruct-AWQ/raw/3559b226e8ce77211e2c1bd7ddfb7686fec4d6dd/config.json" }, { "label": "Pinned base model config", "url": "https://huggingface.co/Qwen/Qwen2.5-3B-Instruct/raw/aa8e72537993ba99e69dfaafa59ed015b17504d1/config.json" }, { "label": "AWQ safetensors header audit", "url": "https://huggingface.co/Qwen/Qwen2.5-3B-Instruct-AWQ/resolve/3559b226e8ce77211e2c1bd7ddfb7686fec4d6dd/model.safetensors" } ] }, { "id": "Hcompany/Holo-3.1-35B-A3B-GGUF", "name": "Holo-3.1-35B-A3B-GGUF GGUF", "short_name": "Holo 3.1 35B A3B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 191258, "tags": [ "transformers", "gguf", "multimodal", "action", "agent", "pytorch", "computer use", "gui agents", "moe", "image-text-to-text", "conversational", "en", "base_model:Qwen/Qwen3.5-0.8B", "base_model:quantized:Qwen/Qwen3.5-0.8B", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen35moe / qwen3_5_moe_text", "total_params_b": 34.660610688, "active_params_b": 3, "layers": 40, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "GGUF BF16/F32", "weight_bytes_per_param": 2, "resident_weight_gb": 69.376637024, "active_weight_gb": 5.937285632, "kv_alloc_gb_per_1k_tokens": 0.02048, "kv_read_gb_per_1k_tokens": 0.02048, "kv_alloc_fixed_gb": 0.06488064, "kv_read_fixed_gb": 0.06488064, "expert_param_b": 0.12582912, "fixed_param_b": 1.962009856, "shared_expert_param_b": 0, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, model card, non-GGUF Holo config/API metadata, Qwen base config, linked GGUF file sizes, and direct selected BF16 GGUF tensor-index range read. The authoritative bounds profile uses main.bf16.gguf because it matches API gguf.totalFileSize: 69.376637024 GB resident and 5.937285632 GB ordinary single-session swept traffic before KV/state reads. The audited profile supersedes the generated dense 35B, over-broad full-KV placeholder." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Hcompany/Holo-3.1-35B-A3B-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Hcompany/Holo-3.1-35B-A3B-GGUF" }, { "label": "Non-GGUF Holo config", "url": "https://huggingface.co/Hcompany/Holo-3.1-35B-A3B/raw/2bdb92851a8cd9d72cdd891fdf38cfcc7fefae2c/config.json" }, { "label": "Qwen base config", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/995ad96eacd98c81ed38be0c5b274b04031597b0/config.json" }, { "label": "Selected GGUF header audit", "url": "https://huggingface.co/Hcompany/Holo-3.1-35B-A3B-GGUF/resolve/95dd5447a6f101de041c7f03821dafac2bac19a4/main.bf16.gguf" } ] }, { "id": "ReadyArt/Dark-Scarlett-v0.3-26B-A4B-GGUF", "name": "Dark-Scarlett-v0.3-26B-A4B-GGUF GGUF IQ4_XS", "short_name": "Dark Scarlett v0.3 26B A4B", "license": "apache-2.0", "hf_pipeline_tag": "unknown", "hf_downloads": 144312, "tags": [ "gguf", "gemma-4", "roleplay", "conversational", "instruct", "apache-2.0", "nsfw", "adult-content", "unaligned", "mature", "explicit", "erp", "base_model:google/gemma-4-26B-A4B-it", "base_model:finetune:google/gemma-4-26B-A4B-it", "license:apache-2.0", "endpoints_compatible", "region:us", "imatrix" ], "architecture": { "type": "MoE transformer", "detail": "gemma4_text", "total_params_b": 25.233142046, "active_params_b": 3.8, "layers": 30, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "GGUF i1-IQ4_XS", "weight_bytes_per_param": 0.5515653344568819, "resident_weight_gb": 13.917726432, "active_weight_gb": 2.304475008, "kv_alloc_gb_per_1k_tokens": 0.20992, "kv_read_gb_per_1k_tokens": 0.20992, "notes": "Manually audited on 2026-07-07 from HF API commit b076da94ae47dd6f270dad74caa5094d6e505266, HF CLI model info, model card metadata, Google base config, linked GGUF file sizes, and a GGUF tensor-index range read. The authoritative bounds profile uses the selected Dark-Scarlett-v0.3-26B-A4B-i1-IQ4_XS.gguf artifact: 13.917726432 GB resident, 1.531313088 GB fixed traffic including tied token_embd/output projection, 0.09664524 GB per routed expert, 8 routed experts per token, 128 routed experts, and layered Gemma 4 hybrid KV. Legacy per-1k KV fields show the short-context unsaturated coefficient; long-context calculation must use the profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/ReadyArt/Dark-Scarlett-v0.3-26B-A4B-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/ReadyArt/Dark-Scarlett-v0.3-26B-A4B-GGUF" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/ReadyArt/Dark-Scarlett-v0.3-26B-A4B-GGUF/resolve/b076da94ae47dd6f270dad74caa5094d6e505266/Dark-Scarlett-v0.3-26B-A4B-i1-IQ4_XS.gguf" } ] }, { "id": "MaziyarPanahi/Meta-Llama-3.1-405B-Instruct-GGUF", "name": "Meta-Llama-3.1-405B-Instruct-GGUF GGUF Q2_K split", "short_name": "Meta Llama 3.1 405B Instruct", "license": "llama3.1", "hf_pipeline_tag": "text-generation", "hf_downloads": 108344, "tags": [ "gguf", "quantized", "2-bit", "3-bit", "GGUF", "text-generation", "en", "de", "fr", "it", "pt", "hi", "es", "th", "base_model:meta-llama/Llama-3.1-405B-Instruct", "base_model:quantized:meta-llama/Llama-3.1-405B-Instruct", "license:llama3.1", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "llama", "total_params_b": 410.081247232, "active_params_b": 410.081247232, "layers": 126, "attention_heads": 128, "kv_heads": 16, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q2_K split", "weight_bytes_per_param": 0.3687208409470545, "resident_weight_gb": 151.205502336, "active_weight_gb": 150.508109824, "kv_alloc_gb_per_1k_tokens": 1.032192, "kv_read_gb_per_1k_tokens": 1.032192, "notes": "Manually audited on 2026-07-07 from HF API commit 85b9bd67025a4337e9694ec0edaf46437fe6283b, HF CLI model info, model card metadata, linked split-GGUF file sizes, and GGUF tensor-index range reads across all nine selected Q2_K shards. The authoritative bounds profile uses the selected 9-part Q2_K split artifact: 151.205502336 GB resident, 150.508109824 GB swept text traffic excluding resident-only token_embd.weight, 0.697392512 GB resident-only token/header overhead, and full-context FP16 Llama KV." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/Meta-Llama-3.1-405B-Instruct-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/meta-llama/Llama-3.1-405B-Instruct/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/meta-llama/Llama-3.1-405B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/Meta-Llama-3.1-405B-Instruct-GGUF" }, { "label": "Selected Q2_K split shard", "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" } ] }, { "id": "MaziyarPanahi/DeepSeek-V3-0324-GGUF", "name": "DeepSeek-V3-0324-GGUF GGUF IQ1_M split", "short_name": "DeepSeek V3 0324", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 108352, "tags": [ "gguf", "quantized", "2-bit", "GGUF", "text-generation", "base_model:deepseek-ai/DeepSeek-V3-0324", "base_model:quantized:deepseek-ai/DeepSeek-V3-0324", "license:mit", "region:us", "imatrix", "conversational" ], "architecture": { "type": "MLA MoE transformer", "detail": "deepseek2 / deepseek_v3", "total_params_b": 671.0264192, "active_params_b": 37, "layers": 61, "attention_heads": 128, "kv_heads": 128, "head_dim": 128, "max_context_tokens": 163840, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "kv_lora_rank": 512, "qk_rope_head_dim": 64, "qk_nope_head_dim": 128, "v_head_dim": 128 }, "adapter": { "kind": "moe", "weight_precision": "GGUF IQ1_M split", "weight_bytes_per_param": 0.22187331745521832, "resident_weight_gb": 148.882857728, "active_weight_gb": 9.204629504, "kv_alloc_gb_per_1k_tokens": 0.070272, "kv_read_gb_per_1k_tokens": 0.070272, "notes": "Manually audited on 2026-07-07 from HF API commit d77054856dd5bb226f066fe76063a3691ffc3cfd, HF CLI model info, model card metadata, official DeepSeek V3 0324 serving adapter evidence, linked split-GGUF file sizes, and GGUF tensor-index range reads across all nine selected IQ1_M shards. The authoritative bounds profile uses the selected 9-part IQ1_M split artifact: 148.882857728 GB resident, 4.708859904 GB fixed traffic excluding resident-only token_embd.weight, 0.5619712 GB per routed expert, 8 routed experts per token, 256 routed experts, and official absorb-MLA BF16 state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MaziyarPanahi/DeepSeek-V3-0324-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3-0324/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3-0324" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MaziyarPanahi/DeepSeek-V3-0324-GGUF" }, { "label": "Selected IQ1_M split shard", "url": "https://huggingface.co/MaziyarPanahi/DeepSeek-V3-0324-GGUF/resolve/d77054856dd5bb226f066fe76063a3691ffc3cfd/DeepSeek-V3-0324.IQ1_M.gguf-00001-of-00009.gguf" } ] }, { "id": "unsloth/Qwen3.5-35B-A3B-GGUF", "name": "Qwen3.5-35B-A3B-GGUF GGUF", "short_name": "Qwen3.5 35B A3B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 150052, "tags": [ "gguf", "qwen3_5_moe", "unsloth", "image-text-to-text", "base_model:Qwen/Qwen3.5-35B-A3B", "base_model:quantized:Qwen/Qwen3.5-35B-A3B", "license:apache-2.0", "endpoints_compatible", "region:us", "imatrix", "conversational" ], "architecture": { "type": "MoE transformer", "detail": "qwen35moe / qwen3_5_moe_text", "total_params_b": 34.660610688, "active_params_b": 3, "layers": 40, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "GGUF MXFP4_MOE", "weight_bytes_per_param": 0.6228291563101036, "resident_weight_gb": 21.587638912, "active_weight_gb": 2.717223424, "kv_alloc_gb_per_1k_tokens": 0.02048, "kv_read_gb_per_1k_tokens": 0.02048, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, served repo config, selected linked-object HEAD checks, direct GGUF header range read, and ggml MXFP4 type definitions. The selected API artifact is Qwen3.5-35B-A3B-MXFP4_MOE.gguf at 21.587638912 GB resident; ordinary single-session swept traffic is 2.717223424 GB before KV/state reads. The audited profile adds the fixed DeltaNet state charge that these legacy per-1K KV fields cannot fully express." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/Qwen3.5-35B-A3B-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/unsloth/Qwen3.5-35B-A3B-GGUF/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3.5-35B-A3B-GGUF" }, { "label": "Selected GGUF header audit", "url": "https://huggingface.co/unsloth/Qwen3.5-35B-A3B-GGUF/resolve/bc014a17be43adabd7066b7a86075ff935c6a4e2/Qwen3.5-35B-A3B-MXFP4_MOE.gguf" }, { "label": "GGUF MXFP4 type definitions", "url": "https://github.com/ggml-org/ggml/blob/master/src/ggml-common.h" } ] }, { "id": "openai/gpt-oss-safeguard-20b", "name": "gpt-oss-safeguard-20b MXFP4", "short_name": "gpt oss safeguard 20b", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 151082, "tags": [ "transformers", "safetensors", "gpt_oss", "text-generation", "vllm", "conversational", "arxiv:2508.10925", "base_model:openai/gpt-oss-20b", "base_model:finetune:openai/gpt-oss-20b", "license:apache-2.0", "endpoints_compatible", "8-bit", "mxfp4", "deploy:azure", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "gpt_oss (GptOssForCausalLM)", "total_params_b": 21.511953984, "active_params_b": 3.6, "layers": 24, "attention_heads": 64, "kv_heads": 8, "head_dim": 64, "max_context_tokens": 131072, "routed_experts": 32, "routed_experts_per_token": 4, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "MXFP4", "weight_bytes_per_param": 0.5, "resident_weight_gb": 13.761264768, "active_weight_gb": 3.708083328, "kv_alloc_gb_per_1k_tokens": 0.02460745728, "kv_read_gb_per_1k_tokens": 0.024674304, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, model card, served config, official gpt-oss-20b profile comparison, safetensors index metadata, linked shard HEAD checks, and direct safetensors header range reads. The authoritative profile records 13.761264768 GB resident MXFP4/BF16 tensor payload, 3.708083328 GB single-token ordinary text-decode traffic, and layered BF16 KV with 12 full-context layers plus 12 sliding-window layers." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/openai/gpt-oss-safeguard-20b" }, { "label": "Model config", "url": "https://huggingface.co/openai/gpt-oss-safeguard-20b/raw/8a11e17b25c973a24099d4016bf2e17dd7ec1574/config.json" }, { "label": "Safetensors index", "url": "https://huggingface.co/openai/gpt-oss-safeguard-20b/raw/8a11e17b25c973a24099d4016bf2e17dd7ec1574/model.safetensors.index.json" }, { "label": "Model card", "url": "https://huggingface.co/openai/gpt-oss-safeguard-20b/raw/8a11e17b25c973a24099d4016bf2e17dd7ec1574/README.md" }, { "label": "Base model card", "url": "https://huggingface.co/openai/gpt-oss-20b" } ] }, { "id": "nvidia/DeepSeek-V4-Pro-NVFP4", "name": "DeepSeek-V4-Pro-NVFP4 NVFP4", "short_name": "DeepSeek V4 Pro NVFP4", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 152280, "tags": [ "Model Optimizer", "safetensors", "deepseek_v4", "nvidia", "ModelOpt", "DeepSeekV4", "quantized", "NVFP4", "nvfp4", "text-generation", "base_model:deepseek-ai/DeepSeek-V4-Pro", "base_model:quantized:deepseek-ai/DeepSeek-V4-Pro", "license:mit", "8-bit", "fp8", "region:us" ], "architecture": { "type": "ModelOpt NVFP4 MoE transformer", "detail": "deepseek_v4 CSA/HCA MoE (DeepseekV4ForCausalLM)", "total_params_b": 1600, "active_params_b": 49, "layers": 61, "attention_heads": 128, "kv_heads": 1, "head_dim": 512, "max_context_tokens": 1048576, "routed_experts": 384, "routed_experts_per_token": 6, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "ModelOpt NVFP4 experts", "weight_bytes_per_param": 0.5, "resident_weight_gb": 913.061485688, "active_weight_gb": 40.440334856, "kv_alloc_gb_per_1k_tokens": 0.009848, "kv_read_gb_per_1k_tokens": 0.002168, "kv_alloc_fixed_gb": 0.02670592, "kv_read_fixed_gb": 0.039452672, "expert_param_b": 2.266695096, "shared_expert_param_b": 0, "fixed_param_b": 26.84016428, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, NVIDIA model card, pinned served config, hf_quant_config, official DeepSeek V4 Pro profile/config/code comparison, linked-object metadata, safetensors index, and direct range-read shard headers. Catalog parameters use the model-card logical 1.6T total / 49B active values. Bound weights use NVIDIA-specific stored bytes: 913.061485688 GB resident tensor payload, 26.840164280 GB fixed ordinary traffic, 2.266695096 GB per routed expert index, and 15.809945772 GB resident-only embed/MTP bytes. KV/state traffic reuses the audited DeepSeek V4 Pro CSA/HCA BF16 path because the ModelOpt sidecar records kv_cache_quant_algo null." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/DeepSeek-V4-Pro-NVFP4" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/DeepSeek-V4-Pro-NVFP4" }, { "label": "Pinned model config", "url": "https://huggingface.co/nvidia/DeepSeek-V4-Pro-NVFP4/raw/1449d1e641023406daf6b432361486c768aad740/config.json" }, { "label": "Pinned ModelOpt quantization config", "url": "https://huggingface.co/nvidia/DeepSeek-V4-Pro-NVFP4/raw/1449d1e641023406daf6b432361486c768aad740/hf_quant_config.json" }, { "label": "Safetensors index and shard-header audit", "url": "https://huggingface.co/nvidia/DeepSeek-V4-Pro-NVFP4/resolve/1449d1e641023406daf6b432361486c768aad740/model.safetensors.index.json" }, { "label": "NVIDIA DeepSeek V4 Pro inference review", "url": "https://huggingface.co/nvidia/DeepSeek-V4-Pro-NVFP4/raw/1449d1e641023406daf6b432361486c768aad740/inference/model.py" }, { "label": "Base model card", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro" } ] }, { "id": "TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ", "name": "TinyLlama-1.1B-Chat-v1.0-GPTQ 4-bit", "short_name": "TinyLlama 1.1B Chat v1.0", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 132526, "tags": [ "transformers", "safetensors", "llama", "text-generation", "conversational", "en", "dataset:cerebras/SlimPajama-627B", "dataset:bigcode/starcoderdata", "dataset:OpenAssistant/oasst_top1_2023-08-25", "base_model:TinyLlama/TinyLlama-1.1B-Chat-v1.0", "base_model:quantized:TinyLlama/TinyLlama-1.1B-Chat-v1.0", "license:apache-2.0", "text-generation-inference", "4-bit", "gptq", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "llama (LlamaForCausalLM)", "total_params_b": 1.100442624, "active_params_b": 1.034906624, "layers": 22, "attention_heads": 32, "kv_heads": 4, "head_dim": 64, "max_context_tokens": 2048, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GPTQ 4-bit group 128 act-order", "weight_bytes_per_param": 0.697954963983656, "resident_weight_gb": 0.768059392, "active_weight_gb": 0.636987392, "kv_alloc_gb_per_1k_tokens": 0.022528, "kv_read_gb_per_1k_tokens": 0.022528, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, pinned model card, pinned GPTQ config, quantize_config, pinned TinyLlama v1.0 base config comparison, and direct safetensors header range reads. Bound weights separate the exact 0.768059392 GB resident GPTQ package from the 0.636987392 GB ordinary text-decode swept traffic; the audited profile supersedes the generated ideal 4-bit estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ" }, { "label": "Model config", "url": "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ/raw/9d4580af0f21bccafd762dcc50d0c7bac6273584/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v1.0" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ" }, { "label": "Quantize config", "url": "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ/raw/9d4580af0f21bccafd762dcc50d0c7bac6273584/quantize_config.json" }, { "label": "Pinned base config", "url": "https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v1.0/raw/fe8a4ea1ffedaf415f4da2f062534de366a451e6/config.json" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ/resolve/9d4580af0f21bccafd762dcc50d0c7bac6273584/model.safetensors" } ] }, { "id": "mlx-community/Devstral-Small-2-24B-Instruct-2512-4bit", "name": "Devstral-Small-2-24B-Instruct-2512-4bit 4-bit", "short_name": "Devstral Small 2 24B Instruct 2512 4bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 129826, "tags": [ "mlx", "safetensors", "mistral3", "mistral-common", "4-bit", "image-text-to-text", "conversational", "base_model:mistralai/Devstral-Small-2-24B-Instruct-2512", "base_model:quantized:mistralai/Devstral-Small-2-24B-Instruct-2512", "license:apache-2.0", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "mistral3 / ministral3 text (MLX 4-bit affine)", "total_params_b": 24.01136128, "active_params_b": 22.90131456, "layers": 40, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 393216, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "MLX 4-bit affine", "weight_bytes_per_param": 0.6289805423310011, "resident_weight_gb": 15.10267904, "active_weight_gb": 13.84727552, "kv_alloc_gb_per_1k_tokens": 0.16384, "kv_read_gb_per_1k_tokens": 0.16384, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, MLX model card, pinned served config and params, pinned Devstral base config/params comparison, safetensors index metadata, linked shard HEAD checks, and direct safetensors header range reads. The authoritative profile records 15.102679040 GB resident MLX U32/BF16 tensor payload, 13.847275520 GB ordinary text-decode swept traffic, and 1.255403520 GB resident-only input embedding plus vision/projector tensors. The current repo metadata points to mistralai/Devstral-Small-2-24B-Instruct-2512, not the older generated Mistral-Small base row." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/mlx-community/Devstral-Small-2-24B-Instruct-2512-4bit" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/mlx-community/Devstral-Small-2-24B-Instruct-2512-4bit" }, { "label": "Pinned model config", "url": "https://huggingface.co/mlx-community/Devstral-Small-2-24B-Instruct-2512-4bit/raw/97228558fcbdd1adefd39d515155db3a894b37f7/config.json" }, { "label": "Pinned params", "url": "https://huggingface.co/mlx-community/Devstral-Small-2-24B-Instruct-2512-4bit/raw/97228558fcbdd1adefd39d515155db3a894b37f7/params.json" }, { "label": "Safetensors index and shard-header audit", "url": "https://huggingface.co/mlx-community/Devstral-Small-2-24B-Instruct-2512-4bit/raw/97228558fcbdd1adefd39d515155db3a894b37f7/model.safetensors.index.json" }, { "label": "Base model card", "url": "https://huggingface.co/mistralai/Devstral-Small-2-24B-Instruct-2512" }, { "label": "Pinned base config", "url": "https://huggingface.co/mistralai/Devstral-Small-2-24B-Instruct-2512/raw/c599e8e56f3f9110e97f0dc0450ce248e3334d84/config.json" } ] }, { "id": "RedHatAI/gemma-4-12B-it-NVFP4", "name": "gemma-4-12B-it-NVFP4 NVFP4", "short_name": "gemma 4 12B it NVFP4", "license": "unknown", "hf_pipeline_tag": "unknown", "hf_downloads": 260994, "tags": [ "safetensors", "gemma4_unified", "gemma", "nvfp4", "vllm", "compressed-tensors", "GPTQ", "base_model:google/gemma-4-12B-it", "base_model:quantized:google/gemma-4-12B-it", "8-bit", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_unified_text", "total_params_b": 8.1976368, "active_params_b": 7.138623936, "layers": 48, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "compressed-tensors NVFP4 + BF16/F8/U8 side tensors", "weight_bytes_per_param": 1.2520747969707562, "resident_weight_gb": 10.264054432, "active_weight_gb": 8.146028704, "kv_alloc_gb_per_1k_tokens": 0.335872, "kv_read_gb_per_1k_tokens": 0.335872, "notes": "Manual correction on 2026-07-07 from live Hugging Face CLI/API metadata, README, served compressed-tensors config, recipe.yaml, Google base config comparison, linked safetensors object metadata, and a direct safetensors header range read. The authoritative bounds profile uses exact stored bytes: 10.264054432 GB resident, 8.146028704 GB swept ordinary text traffic, 2.118025728 GB resident-only embedding and multimodal traffic, separate lm_head output projection despite tie_word_embeddings true, and layered Gemma 4 BF16 KV because kv_cache_scheme is null. Legacy per-1k KV fields show the short-context unsaturated coefficient; long-context calculation must use the profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/RedHatAI/gemma-4-12B-it-NVFP4" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/RedHatAI/gemma-4-12B-it-NVFP4" }, { "label": "Pinned model config", "url": "https://huggingface.co/RedHatAI/gemma-4-12B-it-NVFP4/raw/a1d2478a9b99cc444bc9f64839609d3a82ca9195/config.json" }, { "label": "Pinned recipe", "url": "https://huggingface.co/RedHatAI/gemma-4-12B-it-NVFP4/raw/a1d2478a9b99cc444bc9f64839609d3a82ca9195/recipe.yaml" }, { "label": "Pinned base config", "url": "https://huggingface.co/google/gemma-4-12B-it/raw/5926caa4ec0cac5cbfadaf4077420520de1d5205/config.json" }, { "label": "Safetensors linked-object and header audit", "url": "https://huggingface.co/RedHatAI/gemma-4-12B-it-NVFP4/resolve/a1d2478a9b99cc444bc9f64839609d3a82ca9195/model.safetensors" } ] }, { "id": "lmstudio-community/gemma-4-31B-it-GGUF", "name": "gemma-4-31B-it-GGUF GGUF", "short_name": "gemma 4 31B it", "license": "apache-2.0", "hf_pipeline_tag": "unknown", "hf_downloads": 115645, "tags": [ "gguf", "base_model:google/gemma-4-31B-it", "base_model:quantized:google/gemma-4-31B-it", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 30.697345596, "active_params_b": 30.697345596, "layers": 60, "attention_heads": 32, "kv_heads": 16, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q4_K_M + Q6_K/F32 tensors", "weight_bytes_per_param": 0.6087517113021983, "resident_weight_gb": 18.687061664, "active_weight_gb": 18.671230848, "kv_alloc_gb_per_1k_tokens": 0.86016, "kv_read_gb_per_1k_tokens": 0.86016, "notes": "Manual correction on 2026-07-07 from live Hugging Face CLI/API metadata, model card, Google Gemma 4 31B IT config, HF CLI file listing, and a direct selected GGUF header range read. The authoritative bounds profile uses exact selected-artifact bytes: 18.687061664 GB resident Q4_K_M file, 18.671230848 GB swept tensor spans, 0.015830816 GB metadata/header/alignment overhead, tied token_embd output projection because no output.weight tensor is stored, and layered Gemma 4 FP16 KV from the selected GGUF metadata. Legacy per-1k KV fields show the short-context unsaturated coefficient; long-context calculation must use the profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/gemma-4-31B-it-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/gemma-4-31B-it-GGUF" }, { "label": "Pinned base config", "url": "https://huggingface.co/google/gemma-4-31B-it/raw/3548789868c5356dbf307c98e6f609007b82b3eb/config.json" }, { "label": "Pinned selected GGUF", "url": "https://huggingface.co/lmstudio-community/gemma-4-31B-it-GGUF/resolve/cc1750d5ecdf85a09a19e2649f343e99cdb19524/gemma-4-31B-it-Q4_K_M.gguf" }, { "label": "Pinned file list", "url": "https://huggingface.co/lmstudio-community/gemma-4-31B-it-GGUF/tree/cc1750d5ecdf85a09a19e2649f343e99cdb19524" } ] }, { "id": "unsloth/Kimi-K2.7-Code-GGUF", "name": "Kimi-K2.7-Code-GGUF GGUF 4-bit", "short_name": "Kimi K2.7 Code", "license": "other", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 291513, "tags": [ "transformers", "gguf", "kimi_k25", "feature-extraction", "compressed-tensors", "unsloth", "image-text-to-text", "custom_code", "base_model:moonshotai/Kimi-K2.7-Code", "base_model:quantized:moonshotai/Kimi-K2.7-Code", "license:other", "endpoints_compatible", "region:us", "imatrix", "conversational" ], "architecture": { "type": "MLA MoE transformer", "detail": "deepseek2 / kimi_k25", "total_params_b": 1026.408232448, "active_params_b": 32, "layers": 61, "attention_heads": 64, "kv_heads": 64, "head_dim": 192, "max_context_tokens": 262144, "routed_experts": 384, "routed_experts_per_token": 8, "shared_experts_per_token": 1, "kv_lora_rank": 512, "qk_rope_head_dim": 64, "qk_nope_head_dim": 128, "v_head_dim": 128 }, "adapter": { "kind": "moe", "weight_precision": "GGUF UD-IQ1_M split", "weight_bytes_per_param": 0.29608995802690685, "resident_weight_gb": 303.909170464, "active_weight_gb": 14.664554496, "kv_alloc_gb_per_1k_tokens": 0.070272, "kv_read_gb_per_1k_tokens": 0.070272, "notes": "Manual correction on 2026-07-07 from live Hugging Face CLI/API metadata, model card, repo config, audited Moonshot Kimi K2.7 Code profile, exact selected split linked-object HEAD checks, and direct GGUF header range reads across all eight UD-IQ1_M shards. The authoritative bounds profile uses exact selected-artifact bytes: 303.909170464 GB resident split files, 303.902185472 GB tensor spans, 0.667587872 GB resident-only token embedding plus metadata/header overhead, 8.524617728 GB fixed ordinary text traffic, 0.767492096 GB per routed expert index, and compact MLA state traffic at 0.070272 GB per 1k context tokens." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/Kimi-K2.7-Code-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/Kimi-K2.7-Code-GGUF" }, { "label": "Pinned model config", "url": "https://huggingface.co/unsloth/Kimi-K2.7-Code-GGUF/raw/46352ca8dc32aa60f9754f5bd3fb778deeb9e430/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/moonshotai/Kimi-K2.7-Code" }, { "label": "Selected UD-IQ1_M split shard", "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" } ] }, { "id": "Hyper-AI/Qwen3.5-9B-fp8", "name": "Qwen3.5-9B-fp8 FP8", "short_name": "Qwen3.5 9B fp8", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 104742, "tags": [ "transformers", "safetensors", "qwen3_5", "image-text-to-text", "vlm", "fp8", "quant", "lightvl", "qwen3.5", "conversational", "base_model:Qwen/Qwen3.5-9B-Base", "base_model:quantized:Qwen/Qwen3.5-9B-Base", "license:apache-2.0", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text (LightVL compressed-tensors FP8)", "total_params_b": 9.653104368, "active_params_b": 7.936684544, "layers": 32, "attention_heads": 16, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "hybrid", "weight_precision": "LightVL compressed-tensors FP8", "weight_bytes_per_param": 1.2583921160397424, "resident_weight_gb": 12.147390432, "active_weight_gb": 8.957726208, "kv_alloc_gb_per_1k_tokens": 0.032768, "kv_read_gb_per_1k_tokens": 0.032768, "kv_alloc_fixed_gb": 0.051904512, "kv_read_fixed_gb": 0.051904512, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, pinned LightVL compressed-tensors config, Qwen3.5 9B Base profile comparison, direct safetensors shard-header range reads, and the Transformers qwen3_5 implementation. The authoritative bounds profile uses exact stored bytes: 12.147390432 GB resident, 8.957726208 GB ordinary text-decode swept traffic, 3.189664224 GB resident-only input embedding/vision/MTP tensors, 8 full-attention layers, and a fixed DeltaNet recurrent-state charge of 0.051904512 GB per session/token." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Hyper-AI/Qwen3.5-9B-fp8" }, { "label": "Pinned Hugging Face API metadata", "url": "https://huggingface.co/api/models/Hyper-AI/Qwen3.5-9B-fp8" }, { "label": "Pinned model config", "url": "https://huggingface.co/Hyper-AI/Qwen3.5-9B-fp8/raw/83d256364f0daeacbcb0ce02c39d3a3101752477/config.json" }, { "label": "Pinned BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3.5-9B-Base/raw/68c46c4b3498877f3ef123c856ecfde50c39f404/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Hyper-AI/Qwen3.5-9B-fp8/raw/83d256364f0daeacbcb0ce02c39d3a3101752477/model.safetensors.index.json" }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py" } ] }, { "id": "unsloth/gemma-4-E2B-it-qat-GGUF", "name": "gemma-4-E2B-it-qat-GGUF GGUF UD-Q2_K_XL", "short_name": "gemma 4 E2B it qat", "license": "apache-2.0", "hf_pipeline_tag": "any-to-any", "hf_downloads": 135267, "tags": [ "transformers", "gguf", "gemma4", "image-text-to-text", "unsloth", "gemma", "google", "any-to-any", "base_model:google/gemma-4-E2B-it-qat-q4_0-unquantized", "base_model:quantized:google/gemma-4-E2B-it-qat-q4_0-unquantized", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text (QAT UD-Q2_K_XL GGUF)", "total_params_b": 4.628569635, "active_params_b": 2.279759395, "layers": 35, "attention_heads": 8, "kv_heads": 1, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF UD-Q2_K_XL QAT", "weight_bytes_per_param": 0.47232405265520006, "resident_weight_gb": 2.186184768, "active_weight_gb": 0.849165408, "kv_alloc_gb_per_1k_tokens": 0.006144, "kv_read_gb_per_1k_tokens": 0.014336, "kv_alloc_fixed_gb": 0.006291456, "kv_read_fixed_gb": 0.014680064, "notes": "Audited profile-backed row from live Hugging Face metadata on 2026-07-07; downloads=135267; selected artifact gemma-4-E2B-it-qat-UD-Q2_K_XL.gguf; data quality: API metadata, model card metadata, repo config, QAT source config, Gemma 4 E2B QAT profile comparison, GGUF HEAD checks, Transformers PLE documentation, and direct GGUF header tensor index. Resident weight uses linked file size; active weight excludes the resident-only per-layer token embedding lookup table. Legacy adapter fields approximate the shared-KV rule: full-context traffic is token-proportional, while the sliding-window component is capped at 512 tokens; the self-contained model profile is authoritative for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/gemma-4-E2B-it-qat-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/gemma-4-E2B-it-qat-GGUF" }, { "label": "Selected GGUF 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" }, { "label": "Pinned repo config", "url": "https://huggingface.co/unsloth/gemma-4-E2B-it-qat-GGUF/raw/db01ae3ceeca98487bf3569814f832f5023cd48c/config.json" }, { "label": "Pinned QAT source config", "url": "https://huggingface.co/google/gemma-4-E2B-it-qat-q4_0-unquantized/raw/d40c88a04b9e3aae9cc9d6f63389e786bd20fba4/config.json" }, { "label": "Gemma 4 PLE documentation", "url": "https://huggingface.co/docs/transformers/model_doc/gemma4" }, { "label": "QAT source model card", "url": "https://huggingface.co/google/gemma-4-E2B-it-qat-q4_0-unquantized" } ] }, { "id": "unsloth/Qwen3-4B-GGUF", "name": "Qwen3-4B-GGUF GGUF BF16", "short_name": "Qwen3 4B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 131101, "tags": [ "transformers", "gguf", "qwen3", "text-generation", "qwen", "unsloth", "en", "arxiv:2309.00071", "base_model:Qwen/Qwen3-4B", "base_model:quantized:Qwen/Qwen3-4B", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 GGUF selected BF16 artifact", "total_params_b": 4.022468096, "active_params_b": 4.022468096, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF BF16 selected artifact", "weight_bytes_per_param": 2.001578469697824, "resident_weight_gb": 8.051285536, "active_weight_gb": 8.045328384, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Manual correction on 2026-07-07 from current Hugging Face API metadata, pinned model card, pinned repo config, audited BF16 Qwen3 4B base profile, linked-object HEAD checks, and direct GGUF header range read. Legacy fallback values mirror the audited profile: API-selected Qwen3-4B-BF16.gguf, exact 8.051285536 GB resident linked file, 8.045328384 GB ordinary text-decode swept tensor traffic, 0.005957152 GB resident-only GGUF overhead, tied token_embd output projection, and FP16 full-context KV." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/Qwen3-4B-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/unsloth/Qwen3-4B-GGUF/raw/22c9fc8a8c7700b76a1789366280a6a5a1ad1120/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-4B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3-4B-GGUF" }, { "label": "Selected BF16 GGUF artifact", "url": "https://huggingface.co/unsloth/Qwen3-4B-GGUF/resolve/22c9fc8a8c7700b76a1789366280a6a5a1ad1120/Qwen3-4B-BF16.gguf" } ] }, { "id": "unsloth/Qwen3.5-9B-MTP-GGUF", "name": "Qwen3.5-9B-MTP-GGUF GGUF BF16", "short_name": "Qwen3.5 9B MTP", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 124435, "tags": [ "transformers", "gguf", "unsloth", "qwen", "qwen3_5", "image-text-to-text", "base_model:Qwen/Qwen3.5-9B", "base_model:quantized:Qwen/Qwen3.5-9B", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen35 GGUF selected BF16 MTP artifact", "total_params_b": 9.197093888, "active_params_b": 7.936684544, "layers": 32, "attention_heads": 16, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16/F32 GGUF MTP", "weight_bytes_per_param": 2.0014280545746237, "resident_weight_gb": 18.407321728, "active_weight_gb": 15.87549184, "kv_alloc_gb_per_1k_tokens": 0.032768, "kv_read_gb_per_1k_tokens": 0.032768, "kv_alloc_fixed_gb": 0.051904512, "kv_read_fixed_gb": 0.051904512, "notes": "Manual correction on 2026-07-07 from current Hugging Face API metadata, pinned model card, Qwen3.5 base config, missing package-config check, selected linked-object HEAD checks, direct GGUF header/tensor-index range read, and Qwen3.5 runtime review. downloads=124435. The authoritative profile uses the API-selected Qwen3.5-9B-BF16.gguf MTP artifact: exact 18.407321728 GB resident linked file, 15.875491840 GB ordinary non-speculative text-decode swept traffic, 2.531829888 GB resident-only token embedding plus MTP/header bytes, and Qwen3.5 hybrid KV plus fixed DeltaNet state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/Qwen3.5-9B-MTP-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3.5-9B/raw/c202236235762e1c871ad0ccb60c8ee5ba337b9a/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-9B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3.5-9B-MTP-GGUF" }, { "label": "Selected BF16 MTP GGUF artifact", "url": "https://huggingface.co/unsloth/Qwen3.5-9B-MTP-GGUF/resolve/9716a636ee4bddc3fed678220b7a33dd2a4160ae/Qwen3.5-9B-BF16.gguf" }, { "label": "Transformers Qwen3.5 implementation", "url": "https://github.com/huggingface/transformers/blob/36778030cba4bccfcb15366d7ad860ca30f47417/src/transformers/models/qwen3_5/modeling_qwen3_5.py" } ] }, { "id": "google/gemma-4-12B-it-qat-q4_0-unquantized", "name": "gemma-4-12B-it-qat-q4_0-unquantized BF16", "short_name": "gemma 4 12B it qat q4 0 unquantized", "license": "apache-2.0", "hf_pipeline_tag": "any-to-any", "hf_downloads": 218747, "tags": [ "transformers", "safetensors", "gemma4_unified", "image-text-to-text", "any-to-any", "base_model:google/gemma-4-12B-it", "base_model:finetune:google/gemma-4-12B-it", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_unified_text QAT BF16", "total_params_b": 11.959730224, "active_params_b": 11.90735032, "layers": 48, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 23.919460448, "active_weight_gb": 23.81470064, "kv_alloc_gb_per_1k_tokens": 0.393216, "kv_read_gb_per_1k_tokens": 0.393216, "notes": "Manual correction on 2026-07-07 from current Hugging Face API metadata, served QAT config, base instruction config comparison, and direct safetensors header range read. downloads=218747. This is the unquantized BF16 QAT checkpoint, not a 4-bit runtime artifact. The authoritative profile records exact 23.919460448 GB resident tensor bytes, 23.814700640 GB ordinary text-decode swept language bytes, and 0.104759808 GB resident-only audio/vision projection bytes. Legacy per-1k KV fields are display fallbacks; production bounds use the layered profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/gemma-4-12B-it-qat-q4_0-unquantized" }, { "label": "Model config", "url": "https://huggingface.co/google/gemma-4-12B-it-qat-q4_0-unquantized/raw/58540658b6c08edab2ddc1fbde7f28cc9987ced3/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-12B-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/gemma-4-12B-it-qat-q4_0-unquantized" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/google/gemma-4-12B-it-qat-q4_0-unquantized/resolve/58540658b6c08edab2ddc1fbde7f28cc9987ced3/model.safetensors" } ] }, { "id": "nvidia/KVzap-mlp-Llama-3.1-8B-Instruct", "name": "KVzap-mlp-Llama-3.1-8B-Instruct F32", "short_name": "KVzap mlp Llama 3.1 8B", "license": "apache-2.0", "hf_pipeline_tag": "other", "hf_downloads": 140859, "tags": [ "transformers", "safetensors", "kvzap", "nvidia", "pytorch", "other", "dataset:nvidia/Nemotron-Pretraining-Dataset-sample", "arxiv:2506.05345", "arxiv:2601.07891", "arxiv:2505.23416", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "KV cache pruning surrogate", "detail": "kvzap MLP sidecar (KVzapModel)", "total_params_b": 0.067256576, "active_params_b": 0.067256576, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "F32", "weight_bytes_per_param": 4, "resident_weight_gb": 0.269026304, "active_weight_gb": 0.269026304, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manual correction on 2026-07-07 from current HF CLI/API metadata, pinned config, model card/overview, and direct safetensors header range read. downloads=140859. This repo is a KVzap MLP pruning sidecar that maps host LLM hidden states to KV-head pruning scores, not a standalone text generation model. Production comparisons use the fail-closed profile; no ordinary decode KV geometry is inferred." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/KVzap-mlp-Llama-3.1-8B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/nvidia/KVzap-mlp-Llama-3.1-8B-Instruct/raw/fb936ee2cfebac42c82453100145ff17083c7552/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/KVzap-mlp-Llama-3.1-8B-Instruct" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/nvidia/KVzap-mlp-Llama-3.1-8B-Instruct/resolve/fb936ee2cfebac42c82453100145ff17083c7552/model.safetensors" } ] }, { "id": "llmfan46/gemma-4-31B-it-uncensored-heretic-GGUF", "name": "gemma-4-31B-it-uncensored-heretic-GGUF GGUF", "short_name": "gemma 4 31B it uncensored heretic", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 131249, "tags": [ "transformers", "gguf", "heretic", "uncensored", "decensored", "abliterated", "ara", "image-text-to-text", "base_model:llmfan46/gemma-4-31B-it-uncensored-heretic", "base_model:quantized:llmfan46/gemma-4-31B-it-uncensored-heretic", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text", "total_params_b": 30.697345596, "active_params_b": 30.697345596, "layers": 60, "attention_heads": 32, "kv_heads": 16, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16 GGUF", "weight_bytes_per_param": 2, "resident_weight_gb": 61.413188736, "active_weight_gb": 61.397356416, "kv_alloc_gb_per_1k_tokens": 0.86016, "kv_read_gb_per_1k_tokens": 0.86016, "notes": "Manual correction on 2026-07-07 from current Hugging Face 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 selected GGUF header range read. downloads=131249. The authoritative bounds profile uses exact selected-artifact bytes: 61.413188736 GB resident BF16 GGUF file, 61.397356416 GB swept tensor spans, 0.015832320 GB metadata/header/alignment overhead, tied token_embd output projection because no output.weight tensor is stored, and layered Gemma 4 FP16 KV from the selected GGUF metadata. Separate mmproj GGUF file is not included unless explicitly loaded. Legacy per-1k KV fields show the short-context unsaturated coefficient; long-context calculation must use the profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/llmfan46/gemma-4-31B-it-uncensored-heretic-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/llmfan46/gemma-4-31B-it-uncensored-heretic/raw/031ce2063177d066ff637f15ab59a2ca2eccfd19/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/llmfan46/gemma-4-31B-it-uncensored-heretic" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/llmfan46/gemma-4-31B-it-uncensored-heretic-GGUF" }, { "label": "BF16 GGUF header audit", "url": "https://huggingface.co/llmfan46/gemma-4-31B-it-uncensored-heretic-GGUF/resolve/eee61b81461ac75eb920a24ca9e5d420bb66e33d/gemma-4-31B-it-uncensored-heretic-BF16.gguf" } ] }, { "id": "zai-org/GLM-4.5V", "name": "GLM-4.5V F32/BF16", "short_name": "GLM 4.5V", "license": "mit", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 134414, "tags": [ "transformers", "safetensors", "glm4v_moe", "image-text-to-text", "conversational", "zh", "en", "arxiv:2507.01006", "base_model:zai-org/GLM-4.5-Air-Base", "base_model:finetune:zai-org/GLM-4.5-Air-Base", "license:mit", "endpoints_compatible", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "glm4v_moe_text", "total_params_b": 107.71093312, "active_params_b": 12.803372672, "layers": 46, "attention_heads": 96, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 65536, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "F32/BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 215.42187776, "active_weight_gb": 25.606756864, "kv_alloc_gb_per_1k_tokens": 0.188416, "kv_read_gb_per_1k_tokens": 0.188416, "notes": "Manual correction on 2026-07-07 from current HF CLI/API metadata, pinned config, model card, safetensors index, and direct header range reads across all 46 shards. downloads=134414. Exact tensor payloads total 215.421877760 GB, with 212.463000064 GB main text/logit resident bytes and 2.958877696 GB resident-only visual/input-embedding bytes. Ordinary active traffic is fixed 13.149673984 GB plus 8 expected routed expert groups of 1.557135360 GB each, or 25.606756864 GB. Legacy per-1k KV fields model full-context BF16 KV for 46 layers, 8 KV heads, and 128 head dimension." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/zai-org/GLM-4.5V" }, { "label": "Model config", "url": "https://huggingface.co/zai-org/GLM-4.5V/raw/ed47433b37111465ec527affaaddceff371bca04/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/zai-org/GLM-4.5-Air-Base" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/zai-org/GLM-4.5V" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/zai-org/GLM-4.5V/raw/ed47433b37111465ec527affaaddceff371bca04/model.safetensors.index.json" } ] }, { "id": "jeffcookio/Mistral-Small-3.2-24B-Instruct-2506-awq-sym", "name": "Mistral-Small-3.2-24B-Instruct-2506-awq-sym 4-bit", "short_name": "Mistral Small 3.2 24B Instruct 2506 awq sym", "license": "unknown", "hf_pipeline_tag": "unknown", "hf_downloads": 105570, "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", "region:us" ], "architecture": { "type": "Dense multimodal transformer", "detail": "mistral3_text (compressed-tensors INT4)", "total_params_b": 24.01136128, "active_params_b": 22.90131456, "layers": 40, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "compressed-tensors INT4 symmetric", "weight_bytes_per_param": 0.6257600668611488, "resident_weight_gb": 15.02535104, "active_weight_gb": 12.8052576, "kv_alloc_gb_per_1k_tokens": 0.16384, "kv_read_gb_per_1k_tokens": 0.16384, "notes": "Manual correction on 2026-07-07 from current HF CLI/API metadata, pinned config, model card, pinned Mistral base config/params metadata, safetensors index, and direct header range reads across all four shards. downloads=105570. Exact tensor payloads total 15.025351040 GB. Ordinary text/logit traffic is 12.805257600 GB: language_model.model.layers 11.463070080 GB, language_model.model.norm 0.000010240 GB, and language_model.lm_head 1.342177280 GB. Resident-only input embedding, vision tower, and multimodal projector tensors total 2.220093440 GB. True model params exclude BF16 compression scales and I64 weight-shape side tensors; stored byte traffic charges them." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/jeffcookio/Mistral-Small-3.2-24B-Instruct-2506-awq-sym" }, { "label": "Model config", "url": "https://huggingface.co/jeffcookio/Mistral-Small-3.2-24B-Instruct-2506-awq-sym/raw/49ec31ec0975e46a1aeef24186a45f11be3042b8/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/mistralai/Mistral-Small-3.2-24B-Instruct-2506" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/jeffcookio/Mistral-Small-3.2-24B-Instruct-2506-awq-sym" }, { "label": "Base model config", "url": "https://huggingface.co/mistralai/Mistral-Small-3.2-24B-Instruct-2506/raw/95a6d26c4bfb886c58daf9d3f7332c857cb27b43/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/jeffcookio/Mistral-Small-3.2-24B-Instruct-2506-awq-sym/raw/49ec31ec0975e46a1aeef24186a45f11be3042b8/model.safetensors.index.json" } ] }, { "id": "lyf/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-NVFP4", "name": "Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-NVFP4 NVFP4", "short_name": "Qwen3.6 35B A3B Uncensored HauhauCS Aggressive N", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 136193, "tags": [ "transformers", "safetensors", "qwen3_5_moe", "image-text-to-text", "qwen3.6", "nvfp4", "compressed-tensors", "quantized", "vllm", "moe", "multimodal", "blackwell", "rtx-5090", "sm120", "uncensored", "conversational", "base_model:HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive", "base_model:quantized:HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive", "license:apache-2.0", "endpoints_compatible", "8-bit", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_5_moe_text", "total_params_b": 35.1072, "active_params_b": 3, "layers": 40, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "NVFP4", "weight_bytes_per_param": 0.664727181660742, "resident_weight_gb": 23.336710112, "active_weight_gb": 3.873048576, "kv_alloc_gb_per_1k_tokens": 0.01024, "kv_read_gb_per_1k_tokens": 0.01024, "kv_alloc_fixed_gb": 0.06488064, "kv_read_fixed_gb": 0.06488064, "notes": "Manual correction on 2026-07-07 from current HF CLI/API metadata, pinned config, model card, recipe, HauhauCS base metadata, official Qwen base config, and direct header range-read of model.safetensors. downloads=136193. Exact tensor payloads total 23.336710112 GB. Ordinary text resident bytes are 21.426448896 GB and resident-only visual/input bytes are 1.910261216 GB. Ordinary active traffic is fixed 3.306809856 GB plus eight routed expert groups of 0.070779840 GB each, or 3.873048576 GB. Simple KV coefficients record the FP8 full-attention stream; fixed DeltaNet runtime state is recorded separately." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lyf/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-NVFP4" }, { "label": "Model config", "url": "https://huggingface.co/lyf/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-NVFP4/raw/c9eb0a997ecfc688c468146404606ef6df49555f/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lyf/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-NVFP4" }, { "label": "Quantization recipe", "url": "https://huggingface.co/lyf/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-NVFP4/raw/c9eb0a997ecfc688c468146404606ef6df49555f/recipe.yaml" }, { "label": "Qwen3.6 base config", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/995ad96eacd98c81ed38be0c5b274b04031597b0/config.json" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/lyf/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-NVFP4/resolve/c9eb0a997ecfc688c468146404606ef6df49555f/model.safetensors" } ] }, { "id": "Qwen/Qwen2.5-Omni-7B-AWQ", "name": "Qwen2.5-Omni-7B-AWQ 4-bit", "short_name": "Qwen2.5 Omni 7B", "license": "other", "hf_pipeline_tag": "any-to-any", "hf_downloads": 125455, "tags": [ "transformers", "safetensors", "qwen2_5_omni", "text-to-audio", "multimodal", "any-to-any", "en", "arxiv:2503.20215", "license:other", "endpoints_compatible", "4-bit", "awq", "region:us" ], "architecture": { "type": "Multimodal dense transformer", "detail": "qwen2_5_omni Thinker/Talker package; ordinary text decode uses the 28-layer Thinker text decoder", "total_params_b": 10.73222544, "active_params_b": 7.070619136, "layers": 28, "attention_heads": 28, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "AWQ 4-bit + BF16/F16/F32 side tensors", "weight_bytes_per_param": 0.5, "resident_weight_gb": 12.70206784, "active_weight_gb": 4.48075264, "kv_alloc_gb_per_1k_tokens": 0.057344, "kv_read_gb_per_1k_tokens": 0.057344, "notes": "Manually audited on 2026-07-07 from current HF CLI/API metadata, pinned config, model card, safetensors index, and range-read safetensors headers. downloads=125455. The authoritative bounds profile uses exact stored bytes: 12.70206784 GB resident, 4.48075264 GB swept text traffic, 8.2213152 GB resident-only multimodal/input-embedding traffic, and full-context BF16 Thinker text KV." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Omni-7B-AWQ" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-Omni-7B-AWQ/raw/848615c28dfd87b0f624baea4be1f29bec0d1db1/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-Omni-7B-AWQ" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/Qwen/Qwen2.5-Omni-7B-AWQ/raw/848615c28dfd87b0f624baea4be1f29bec0d1db1/model.safetensors.index.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Omni-7B" } ] }, { "id": "nvidia/Nemotron-Labs-Diffusion-8B", "name": "Nemotron-Labs-Diffusion-8B BF16", "short_name": "Nemotron Labs Diffusion 8B", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 125076, "tags": [ "transformers", "safetensors", "nemotron_labs_diffusion", "feature-extraction", "nvidia", "pytorch", "text-generation", "conversational", "custom_code", "license:other", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "nemotron_labs_diffusion tri-mode decoder", "total_params_b": 8.48955392, "active_params_b": 7.952683008, "layers": 34, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 16.97910784, "active_weight_gb": 15.905366016, "kv_alloc_gb_per_1k_tokens": 0.139264, "kv_read_gb_per_1k_tokens": 0.139264, "notes": "Manual correction on 2026-07-07 from live Hugging Face API metadata, pinned config/generation config, model card, custom modeling code, and direct safetensors header range reads. downloads=125076. The core checkpoint has 16.979107840 GB BF16 resident tensors; AR-mode non-embedding swept bytes are 15.905366016 GB. Production bounds are disabled in the model profile because the repo exposes AR, diffusion, and self-speculation modes without a comparison-time generation-mode selector." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-8B" }, { "label": "Model config", "url": "https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-8B/raw/16c67f0560b912e93e0cabb6e0c4f5c3086d95fc/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/Nemotron-Labs-Diffusion-8B" }, { "label": "Custom modeling code", "url": "https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-8B/raw/16c67f0560b912e93e0cabb6e0c4f5c3086d95fc/modeling_nemotron_labs_diffusion.py" }, { "label": "Pinned safetensors header audit", "url": "https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-8B/resolve/16c67f0560b912e93e0cabb6e0c4f5c3086d95fc/model.safetensors" } ] }, { "id": "allenai/OLMoE-1B-7B-0924", "name": "OLMoE-1B-7B-0924 BF16", "short_name": "OLMoE 1B 7B 0924", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 144794, "tags": [ "transformers", "safetensors", "olmoe", "text-generation", "moe", "olmo", "en", "dataset:allenai/OLMoE-mix-0924", "arxiv:2409.02060", "license:apache-2.0", "co2_eq_emissions", "endpoints_compatible", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "olmoe (OlmoeForCausalLM)", "total_params_b": 6.919161856, "active_params_b": 1.178994688, "layers": 16, "attention_heads": 16, "kv_heads": 16, "head_dim": 128, "max_context_tokens": 4096, "routed_experts": 64, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 13.838323712, "active_weight_gb": 2.357989376, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "expert_param_b": 0.100663296, "fixed_param_b": 0.37368832, "notes": "Manually audited on 2026-07-07 from current Hugging Face CLI/API metadata, pinned config, model card, safetensors index, and direct range-read shard headers. downloads=144794. Bound weights use exact stored bytes: 13.838323712 GB resident, 0.747376640 GB fixed ordinary traffic, 0.201326592 GB per routed expert index across 64 experts, 8 routed experts per token, and 0.206045184 GB resident-only input embeddings. The card labels the model 1B active / 7B total and reports 1.3B active in the evaluation table; this row records the exact single-token memory traffic implied by the stored tensors." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/allenai/OLMoE-1B-7B-0924" }, { "label": "Model config", "url": "https://huggingface.co/allenai/OLMoE-1B-7B-0924/raw/6d84c48581ece794365f2b8e9cfb043c68ade9c5/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/allenai/OLMoE-1B-7B-0924" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/allenai/OLMoE-1B-7B-0924/raw/6d84c48581ece794365f2b8e9cfb043c68ade9c5/model.safetensors.index.json" } ] }, { "id": "cyankiwi/Qwen3-Coder-Next-AWQ-4bit", "name": "Qwen3-Coder-Next-AWQ-4bit 4-bit", "short_name": "Qwen3 Coder Next AWQ 4bit", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 118028, "tags": [ "transformers", "safetensors", "qwen3_next", "text-generation", "conversational", "base_model:Qwen/Qwen3-Coder-Next", "base_model:quantized:Qwen/Qwen3-Coder-Next", "license:apache-2.0", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_next (Qwen3NextForCausalLM)", "total_params_b": 79.674391296, "active_params_b": 3.56376448, "layers": 48, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "compressed-tensors AWQ INT4", "weight_bytes_per_param": 0.6051842087737511, "resident_weight_gb": 48.217683456, "active_weight_gb": 4.95699968, "kv_alloc_gb_per_1k_tokens": 0.024576, "kv_read_gb_per_1k_tokens": 0.024576, "kv_alloc_fixed_gb": 0.077856768, "kv_read_fixed_gb": 0.077856768, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, pinned compressed-tensors AWQ config and recipe, official BF16 base config comparison, safetensors index, direct shard-header range reads, and the Transformers qwen3_next implementation. Bound weights use the exact 48.217683456 GB resident package and 4.956999680 GB single-session ordinary text-decode traffic; the audited profile supersedes the generated ideal 4-bit estimate, fixes shared expert count, and adds the fixed DeltaNet runtime state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/cyankiwi/Qwen3-Coder-Next-AWQ-4bit" }, { "label": "Pinned Hugging Face API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3-Coder-Next-AWQ-4bit" }, { "label": "Pinned model config", "url": "https://huggingface.co/cyankiwi/Qwen3-Coder-Next-AWQ-4bit/raw/fa29333988617831d44a4bb68ee3ebde2616346c/config.json" }, { "label": "Pinned compression recipe", "url": "https://huggingface.co/cyankiwi/Qwen3-Coder-Next-AWQ-4bit/raw/fa29333988617831d44a4bb68ee3ebde2616346c/recipe.yaml" }, { "label": "Pinned BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3-Coder-Next/raw/a7fbcb5c0e12d62a448eaa0e260346bf5dcc0feb/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/cyankiwi/Qwen3-Coder-Next-AWQ-4bit/resolve/fa29333988617831d44a4bb68ee3ebde2616346c/model.safetensors.index.json" }, { "label": "Transformers Qwen3-Next implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/qwen3_next/modeling_qwen3_next.py" } ] }, { "id": "sakamakismile/Qwen3.6-27B-NVFP4", "name": "Qwen3.6-27B-NVFP4 NVFP4", "short_name": "Qwen3.6 27B NVFP4", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 169313, "tags": [ "safetensors", "qwen3_5", "image-text-to-text", "base_model:Qwen/Qwen3.6-27B", "base_model:quantized:Qwen/Qwen3.6-27B", "license:apache-2.0", "8-bit", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 27.35672856, "active_params_b": 25.624600064, "layers": 64, "attention_heads": 24, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "NVFP4", "weight_bytes_per_param": 0.5, "resident_weight_gb": 19.709536608, "active_weight_gb": 16.245279616, "kv_alloc_gb_per_1k_tokens": 0.032768, "kv_read_gb_per_1k_tokens": 0.032768, "kv_alloc_fixed_gb": 0.154927104, "kv_read_fixed_gb": 0.154927104, "notes": "Manual correction on 2026-07-07 from current Hugging Face API metadata, model card, pinned compressed-tensors NVFP4 config and recipe, current base config comparison, and direct range-read safetensors header. Bound weights separate the 19.709536608 GB resident NVFP4/BF16/F8/F32 package from the 16.245279616 GB ordinary text-decode swept subset; attention KV follows the model card's production FP8 KV launch and adds fixed DeltaNet runtime state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/sakamakismile/Qwen3.6-27B-NVFP4" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/sakamakismile/Qwen3.6-27B-NVFP4" }, { "label": "Pinned NVFP4 config", "url": "https://huggingface.co/sakamakismile/Qwen3.6-27B-NVFP4/raw/c12989315a26e1cc5d3f8a5d4b96fdd5921adc8c/config.json" }, { "label": "Pinned NVFP4 recipe", "url": "https://huggingface.co/sakamakismile/Qwen3.6-27B-NVFP4/raw/c12989315a26e1cc5d3f8a5d4b96fdd5921adc8c/recipe.yaml" }, { "label": "Base config comparison", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/sakamakismile/Qwen3.6-27B-NVFP4/resolve/c12989315a26e1cc5d3f8a5d4b96fdd5921adc8c/model.safetensors" } ] }, { "id": "zaakirio/gemma-4-12b-it-uncensored-GGUF", "name": "gemma-4-12b-it-uncensored-GGUF GGUF", "short_name": "gemma 4 12b it uncensored", "license": "gemma", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 102022, "tags": [ "gguf", "llama.cpp", "gemma", "gemma4", "heretic", "abliterated", "uncensored", "decensored", "conversational", "multimodal", "audio", "image-text-to-text", "arxiv:2512.13655", "arxiv:2406.11717", "base_model:zaakirio/gemma-4-12b-it-uncensored", "base_model:quantized:zaakirio/gemma-4-12b-it-uncensored", "license:gemma", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_unified_text", "total_params_b": 11.907350576, "active_params_b": 11.907350576, "layers": 48, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q4_K_M", "weight_bytes_per_param": 0.619901271310314, "resident_weight_gb": 7.38138176, "active_weight_gb": 7.365559808, "kv_alloc_gb_per_1k_tokens": 0.335872, "kv_read_gb_per_1k_tokens": 0.335872, "notes": "Manually audited on 2026-07-07 from HF API commit 32880562ac43cb589a85afb864309fdcaf486fae, model card, source config, linked GGUF file sizes, and a GGUF tensor-index range read. The authoritative bounds profile uses the card-recommended gemma-4-12b-it-uncensored-Q4_K_M.gguf artifact: 7.381381760 GB resident, 7.365559808 GB swept text traffic, 0.015821952 GB resident-only GGUF metadata/header bytes, and layered Gemma 4 hybrid KV. API gguf.totalFileSize points to f16, so the selected artifact is intentionally pinned from the card examples. Separate mmproj and MTP sidecars are not included unless explicitly loaded. Legacy per-1k KV fields show the short-context unsaturated coefficient; long-context calculation must use the profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/zaakirio/gemma-4-12b-it-uncensored-GGUF" }, { "label": "Source model config", "url": "https://huggingface.co/zaakirio/gemma-4-12b-it-uncensored/raw/5fdfef5f3aa9dd3c680a22078f2aa97b4a31c919/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/zaakirio/gemma-4-12b-it-uncensored" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/zaakirio/gemma-4-12b-it-uncensored-GGUF" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/zaakirio/gemma-4-12b-it-uncensored-GGUF/resolve/32880562ac43cb589a85afb864309fdcaf486fae/gemma-4-12b-it-uncensored-Q4_K_M.gguf" } ] }, { "id": "google/reformer-crime-and-punishment", "name": "reformer-crime-and-punishment FP32", "short_name": "reformer crime and punishment", "license": "unknown", "hf_pipeline_tag": "text-generation", "hf_downloads": 105506, "tags": [ "transformers", "pytorch", "rust", "reformer", "text-generation", "endpoints_compatible", "region:us" ], "architecture": { "type": "Reformer language model", "detail": "reformer (ReformerModelWithLMHead local/LSH)", "total_params_b": 0.002748544, "active_params_b": 0.002437248, "layers": 6, "attention_heads": 2, "kv_heads": null, "head_dim": 64, "max_context_tokens": 524288, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "FP32", "weight_bytes_per_param": 4, "resident_weight_gb": 0.010994176, "active_weight_gb": 0.009748992, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manual fail-closed correction on 2026-07-07 from live Hugging Face API metadata, pinned config/model card, linked-object HEAD checks, a metadata-only unpickle of the pinned PyTorch state_dict, and Transformers Reformer runtime review. downloads=105506. The PyTorch checkpoint stores 77 FP32 tensors totaling 10.994176 MB / 2.748544M parameters; non-embedding tensors total 9.748992 MB / 2.437248M parameters. Production bounds are disabled in the model profile because Reformer uses cached hidden states plus LSH bucket assignments rather than ordinary per-layer K/V tensors." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/google/reformer-crime-and-punishment" }, { "label": "Model config", "url": "https://huggingface.co/google/reformer-crime-and-punishment/raw/8fcd50a33ea773f3bacf3fecdf193c9ca2c060ac/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/google/reformer-crime-and-punishment" }, { "label": "Pinned PyTorch checkpoint", "url": "https://huggingface.co/google/reformer-crime-and-punishment/resolve/8fcd50a33ea773f3bacf3fecdf193c9ca2c060ac/pytorch_model.bin" }, { "label": "Transformers Reformer runtime", "url": "https://github.com/huggingface/transformers/blob/v4.55.0/src/transformers/models/reformer/modeling_reformer.py" } ] }, { "id": "lmstudio-community/Qwen2.5-Coder-14B-Instruct-MLX-8bit", "name": "Qwen2.5-Coder-14B-Instruct-MLX-8bit 8-bit", "short_name": "Qwen2.5 Coder 14B Instruct MLX 8bit", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 118924, "tags": [ "mlx", "safetensors", "qwen2", "code", "codeqwen", "chat", "qwen", "qwen-coder", "text-generation", "conversational", "en", "base_model:Qwen/Qwen2.5-Coder-14B-Instruct", "base_model:quantized:Qwen/Qwen2.5-Coder-14B-Instruct", "license:apache-2.0", "8-bit", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 14.770033664, "active_params_b": 13.991465984, "layers": 48, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "MLX 8-bit affine", "weight_bytes_per_param": 1.062553362099094, "resident_weight_gb": 15.693948928, "active_weight_gb": 14.866720768, "kv_alloc_gb_per_1k_tokens": 0.196608, "kv_read_gb_per_1k_tokens": 0.196608, "notes": "Manually audited on 2026-07-07 from live Hugging Face API metadata, pinned served MLX config, audited BF16 base profile/config comparison, audited MLX 4-bit sibling profile, model card, safetensors index, and direct three-shard safetensors header reads. downloads=118924. Bound weights use exact tensor spans: 15.693948928 GB resident, 14.866720768 GB ordinary swept text-decode traffic, 0.827228160 GB resident-only input embeddings, and BF16 full-context Qwen2 KV." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/Qwen2.5-Coder-14B-Instruct-MLX-8bit" }, { "label": "Model config", "url": "https://huggingface.co/lmstudio-community/Qwen2.5-Coder-14B-Instruct-MLX-8bit/raw/6fec0bda9a7de3b76be0395d35fa9045cc9d4c68/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-14B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen2.5-Coder-14B-Instruct-MLX-8bit" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/lmstudio-community/Qwen2.5-Coder-14B-Instruct-MLX-8bit/raw/6fec0bda9a7de3b76be0395d35fa9045cc9d4c68/model.safetensors.index.json" } ] }, { "id": "cyankiwi/Qwen3-VL-4B-Instruct-AWQ-4bit", "name": "Qwen3-VL-4B-Instruct-AWQ-4bit 4-bit", "short_name": "Qwen3 VL 4B Instruct AWQ 4bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 116951, "tags": [ "safetensors", "qwen3_vl", "image-text-to-text", "conversational", "arxiv:2505.09388", "arxiv:2502.13923", "arxiv:2409.12191", "arxiv:2308.12966", "base_model:Qwen/Qwen3-VL-4B-Instruct", "base_model:quantized:Qwen/Qwen3-VL-4B-Instruct", "license:apache-2.0", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_vl_text", "total_params_b": 4.940313592, "active_params_b": 4.13600972, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "AWQ 4-bit compressed-tensors", "weight_bytes_per_param": 0.8968370864502805, "resident_weight_gb": 4.430656448, "active_weight_gb": 2.822048704, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Manually audited on 2026-07-07 from live Hugging Face API metadata, pinned served compressed-tensors config, Qwen BF16 base config comparison, recipe.yaml, and direct safetensors header read. downloads=116951. Bound weights use exact tensor spans: 4.430656448 GB resident, 2.822048704 GB ordinary swept text-decode traffic, 1.608607744 GB resident-only visual/input-embedding payload, and BF16 full-context Qwen3-VL KV." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/cyankiwi/Qwen3-VL-4B-Instruct-AWQ-4bit" }, { "label": "Model config", "url": "https://huggingface.co/cyankiwi/Qwen3-VL-4B-Instruct-AWQ-4bit/raw/5ab8b8a020ffd84a86008d4f628d796f1827f182/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3-VL-4B-Instruct-AWQ-4bit" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/cyankiwi/Qwen3-VL-4B-Instruct-AWQ-4bit/resolve/5ab8b8a020ffd84a86008d4f628d796f1827f182/model.safetensors" } ] }, { "id": "allenai/Olmo-3-7B-Instruct", "name": "Olmo-3-7B-Instruct BF16", "short_name": "Olmo 3 7B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 144100, "tags": [ "transformers", "safetensors", "olmo3", "text-generation", "conversational", "en", "dataset:allenai/Dolci-Instruct-RL-7B", "arxiv:2512.13961", "base_model:allenai/Olmo-3-7B-Instruct-DPO", "base_model:finetune:allenai/Olmo-3-7B-Instruct-DPO", "license:apache-2.0", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "olmo3 (Olmo3ForCausalLM)", "total_params_b": 7.298011136, "active_params_b": 7.298011136, "layers": 32, "attention_heads": 32, "kv_heads": 32, "head_dim": 128, "max_context_tokens": 65536, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 14.596022272, "active_weight_gb": 13.774544896, "kv_alloc_gb_per_1k_tokens": 0.524288, "kv_read_gb_per_1k_tokens": 0.524288, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, pinned model card, pinned config, DPO/SFT config comparison, Transformers OLMo3 implementation, linked object metadata, safetensors index, and direct safetensors shard-header range reads. Direct headers found 14.596022272 GB resident BF16 tensor payload, 13.774544896 GB ordinary text-decode swept traffic, and 0.821477376 GB resident-only input embedding. The authoritative profile uses layered KV: 8 full-context layers plus 24 sliding-window layers capped at 4096 tokens." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/allenai/Olmo-3-7B-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/allenai/Olmo-3-7B-Instruct/raw/6e5971d9eba42665f5bd5a0fcf047f299ce1dccc/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/allenai/Olmo-3-7B-Instruct" }, { "label": "Base model card", "url": "https://huggingface.co/allenai/Olmo-3-7B-Instruct-DPO" }, { "label": "Safetensors index and shard-header audit", "url": "https://huggingface.co/allenai/Olmo-3-7B-Instruct/raw/6e5971d9eba42665f5bd5a0fcf047f299ce1dccc/model.safetensors.index.json" } ] }, { "id": "zai-org/chatglm3-6b", "name": "chatglm3-6b FP16", "short_name": "chatglm3 6b", "license": "unknown", "hf_pipeline_tag": "unknown", "hf_downloads": 114420, "tags": [ "transformers", "pytorch", "safetensors", "chatglm", "glm", "thudm", "custom_code", "zh", "en", "arxiv:2103.10360", "arxiv:2210.02414", "arxiv:2406.12793", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "chatglm (ChatGLMModel)", "total_params_b": 6.243584032, "active_params_b": 5.977245728, "layers": 28, "attention_heads": 32, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 8192, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "FP16", "weight_bytes_per_param": 2, "resident_weight_gb": 12.487168064, "active_weight_gb": 11.954491456, "kv_alloc_gb_per_1k_tokens": 0.028672, "kv_read_gb_per_1k_tokens": 0.028672, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, model card, pinned config, remote modeling code, linked object metadata, safetensors index, and direct safetensors shard-header range reads. Direct headers found 12.487168064 GB resident FP16 tensor payload, 11.954491456 GB ordinary text-decode swept traffic, and 0.532676608 GB resident-only input embedding. The authoritative profile uses full-context multi-query KV with 2 KV heads across 28 layers and an 8192-token context." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/zai-org/chatglm3-6b" }, { "label": "Model config", "url": "https://huggingface.co/zai-org/chatglm3-6b/raw/e9e0406d062cdb887444fe5bd546833920abd4ac/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/zai-org/chatglm3-6b" }, { "label": "Remote modeling code", "url": "https://huggingface.co/zai-org/chatglm3-6b/raw/e9e0406d062cdb887444fe5bd546833920abd4ac/modeling_chatglm.py" }, { "label": "Safetensors index and shard-header audit", "url": "https://huggingface.co/zai-org/chatglm3-6b/resolve/e9e0406d062cdb887444fe5bd546833920abd4ac/model.safetensors.index.json" } ] }, { "id": "llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-GGUF", "name": "Qwen3.6-35B-A3B-uncensored-heretic-GGUF GGUF BF16", "short_name": "Qwen3.6 35B A3B uncensored heretic", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 145539, "tags": [ "transformers", "gguf", "qwen3_5_moe", "heretic", "uncensored", "decensored", "abliterated", "image-text-to-text", "base_model:llmfan46/Qwen3.6-35B-A3B-uncensored-heretic", "base_model:quantized:llmfan46/Qwen3.6-35B-A3B-uncensored-heretic", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "MoE transformer", "detail": "qwen35moe / qwen3_5_moe_text", "total_params_b": 34.660610688, "active_params_b": 3, "layers": 40, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "GGUF BF16/F32", "weight_bytes_per_param": 2, "resident_weight_gb": 69.376641152, "active_weight_gb": 5.937285632, "kv_alloc_gb_per_1k_tokens": 0.02048, "kv_read_gb_per_1k_tokens": 0.02048, "kv_alloc_fixed_gb": 0.06488064, "kv_read_fixed_gb": 0.06488064, "expert_param_b": 0.12582912, "fixed_param_b": 1.962009856, "shared_expert_param_b": 0, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, model card, source Heretic config/API metadata, Qwen base config, linked GGUF file sizes, and direct selected BF16 GGUF tensor-index range read. The authoritative bounds profile uses Qwen3.6-35B-A3B-uncensored-heretic-BF16.gguf because it matches API gguf.totalFileSize: 69.376641152 GB resident and 5.937285632 GB ordinary single-session swept traffic before KV/state reads. The audited profile supersedes the generated stale download count, zero shared-expert placeholder, rounded 35.1072B total, and over-broad full-KV estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-GGUF" }, { "label": "Source Heretic config", "url": "https://huggingface.co/llmfan46/Qwen3.6-35B-A3B-uncensored-heretic/raw/dbfd9eb0cdc7c33fc970b06429f6e043b6851190/config.json" }, { "label": "Source Heretic model card", "url": "https://huggingface.co/llmfan46/Qwen3.6-35B-A3B-uncensored-heretic" }, { "label": "Qwen base config", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/995ad96eacd98c81ed38be0c5b274b04031597b0/config.json" }, { "label": "Selected GGUF header audit", "url": "https://huggingface.co/llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-GGUF/resolve/f70be2db155a4192a59c559ece01572f3cd508ab/Qwen3.6-35B-A3B-uncensored-heretic-BF16.gguf" } ] }, { "id": "MiniMaxAI/MiniMax-M2", "name": "MiniMax-M2 FP8", "short_name": "MiniMax M2", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 113607, "tags": [ "transformers", "safetensors", "minimax_m2", "text-generation", "conversational", "custom_code", "arxiv:2504.07164", "arxiv:2509.06501", "arxiv:2509.13160", "license:other", "eval-results", "endpoints_compatible", "fp8", "deploy:azure", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "minimax_m2 (MiniMaxM2ForCausalLM)", "total_params_b": 228.703644928, "active_params_b": 10.41655168, "layers": 62, "attention_heads": 48, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 196608, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "FP8", "weight_bytes_per_param": 1, "resident_weight_gb": 230.12163072, "active_weight_gb": 11.180086272, "kv_alloc_gb_per_1k_tokens": 0.253952, "kv_read_gb_per_1k_tokens": 0.253952, "notes": "Manually corrected on 2026-07-07 from current Hugging Face API metadata, served config, custom configuration/model code, safetensors index, and direct safetensors header range reads. Active traffic is fixed ordinary text traffic plus 8 routed expert groups; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M2" }, { "label": "Model config", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M2/raw/757303d492a50514c312788b5247a4f696a4c6a3/config.json" }, { "label": "Custom model code", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M2/raw/757303d492a50514c312788b5247a4f696a4c6a3/modeling_minimax_m2.py" }, { "label": "Safetensors index", "url": "https://huggingface.co/MiniMaxAI/MiniMax-M2/raw/757303d492a50514c312788b5247a4f696a4c6a3/model.safetensors.index.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/MiniMaxAI/MiniMax-M2" } ] }, { "id": "Qwen/Qwen1.5-7B", "name": "Qwen1.5-7B BF16", "short_name": "Qwen1.5 7B", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 111122, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "pretrained", "conversational", "en", "arxiv:2309.16609", "license:other", "text-generation-inference", "endpoints_compatible", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 7.721324544, "active_params_b": 7.098994688, "layers": 32, "attention_heads": 32, "kv_heads": 32, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 15.442649088, "active_weight_gb": 14.197989376, "kv_alloc_gb_per_1k_tokens": 0.524288, "kv_read_gb_per_1k_tokens": 0.524288, "notes": "Manually corrected on 2026-07-07 from current Hugging Face API metadata, model card, pinned config, safetensors index, and direct safetensors shard-header range reads. The active text-decode traffic excludes only the resident input embedding lookup and includes the separate lm_head.weight output projection." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen1.5-7B" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen1.5-7B/raw/831096e3a59a0789a541415da25ef195ceb802fe/config.json" }, { "label": "Safetensors index", "url": "https://huggingface.co/Qwen/Qwen1.5-7B/raw/831096e3a59a0789a541415da25ef195ceb802fe/model.safetensors.index.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen1.5-7B" } ] }, { "id": "unsloth/Qwen3-8B-GGUF", "name": "Qwen3-8B-GGUF GGUF BF16", "short_name": "Qwen3 8B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 110310, "tags": [ "transformers", "gguf", "qwen3", "text-generation", "qwen", "unsloth", "en", "arxiv:2309.00071", "base_model:Qwen/Qwen3-8B", "base_model:quantized:Qwen/Qwen3-8B", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 GGUF selected BF16 artifact", "total_params_b": 8.19073536, "active_params_b": 7.568405504, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF BF16 selected artifact", "weight_bytes_per_param": 2.000802573115975, "resident_weight_gb": 16.388044384, "active_weight_gb": 15.137427456, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Manual correction on 2026-07-07 from current Hugging Face API metadata, pinned model card, served config, audited BF16 Qwen3 8B base profile, linked-object range checks, and direct GGUF header range read. Legacy fallback values mirror the audited profile: API-selected Qwen3-8B-BF16.gguf, exact 16.388044384 GB resident linked file, 15.137427456 GB ordinary text-decode swept tensor traffic, 1.250616928 GB resident-only input embedding plus GGUF overhead, and FP16 full-context KV. The authoritative profile supersedes the generated full-resident GGUF estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/Qwen3-8B-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/unsloth/Qwen3-8B-GGUF/raw/a6adef130ffb23ddaf1a62fec9dced968c9bc482/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-8B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3-8B-GGUF" }, { "label": "Selected BF16 GGUF artifact", "url": "https://huggingface.co/unsloth/Qwen3-8B-GGUF/resolve/a6adef130ffb23ddaf1a62fec9dced968c9bc482/Qwen3-8B-BF16.gguf" } ] }, { "id": "allenai/OLMoE-1B-7B-0125-Instruct", "name": "OLMoE-1B-7B-0125-Instruct BF16", "short_name": "OLMoE 1B 7B 0125", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 125412, "tags": [ "transformers", "safetensors", "olmoe", "text-generation", "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", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "olmoe (OlmoeForCausalLM)", "total_params_b": 6.919161856, "active_params_b": 1.178994688, "layers": 16, "attention_heads": 16, "kv_heads": 16, "head_dim": 128, "max_context_tokens": 4096, "routed_experts": 64, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "BF16", "weight_bytes_per_param": 2, "resident_weight_gb": 13.838323712, "active_weight_gb": 2.357989376, "kv_alloc_gb_per_1k_tokens": 0.131072, "kv_read_gb_per_1k_tokens": 0.131072, "expert_param_b": 0.100663296, "fixed_param_b": 0.37368832, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, pinned model card, pinned config, safetensors index, direct range-read shard headers, and the audited OLMoE 1B-7B family geometry. downloads=125412. Bound weights use exact stored bytes: 13.838323712 GB resident, 0.747376640 GB fixed ordinary traffic, 0.201326592 GB per routed expert index across 64 experts, 8 routed experts per token, and 0.206045184 GB resident-only input embeddings. The config default records use_cache false, but the architecture remains full-attention OLMoE; this row records cached-decode memory traffic for ordinary serving." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/allenai/OLMoE-1B-7B-0125-Instruct" }, { "label": "Model config", "url": "https://huggingface.co/allenai/OLMoE-1B-7B-0125-Instruct/raw/b89a7c4bc24fb9e55ce2543c9458ce0ca5c4650e/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/allenai/OLMoE-1B-7B-0125-DPO" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/allenai/OLMoE-1B-7B-0125-Instruct" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/allenai/OLMoE-1B-7B-0125-Instruct/raw/b89a7c4bc24fb9e55ce2543c9458ce0ca5c4650e/model.safetensors.index.json" } ] }, { "id": "microsoft/kosmos-2.5", "name": "kosmos-2.5 F32", "short_name": "kosmos 2.5", "license": "mit", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 100827, "tags": [ "transformers", "safetensors", "kosmos-2.5", "image-text-to-text", "en", "arxiv:2309.11419", "license:mit", "endpoints_compatible", "region:us" ], "architecture": { "type": "Multimodal document transformer", "detail": "kosmos-2.5 image-text-to-text (Kosmos2_5ForConditionalGeneration)", "total_params_b": 1.374646272, "active_params_b": 1.374646272, "layers": 24, "attention_heads": 16, "kv_heads": null, "head_dim": null, "max_context_tokens": 4096, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "F32", "weight_bytes_per_param": 4, "resident_weight_gb": 5.498585088, "active_weight_gb": 5.498585088, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manual unsupported correction on 2026-07-07 from current Hugging Face API metadata, pinned model card, pinned config, safetensors index, and direct range-read shard headers. downloads=100827. The stored F32 tensor payload is 5.498585088 GB: 3.387537408 GB text_model, 2.051248128 GB vision_model, and 0.059799552 GB image_to_text_projection. The authoritative profile fails closed because Bounds Engine v1 does not model Kosmos-style image-conditioned OCR/markdown generation with flattened image patches and latent queries." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/microsoft/kosmos-2.5" }, { "label": "Model config", "url": "https://huggingface.co/microsoft/kosmos-2.5/raw/ec3c8051b697166514a31d646cfa36d6ef4c93d7/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/microsoft/kosmos-2.5" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/microsoft/kosmos-2.5/raw/ec3c8051b697166514a31d646cfa36d6ef4c93d7/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen3-VL-235B-A22B-Instruct-FP8", "name": "Qwen3-VL-235B-A22B-Instruct-FP8 FP8", "short_name": "Qwen3 VL 235B A22B Instruct FP8", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 113778, "tags": [ "transformers", "safetensors", "qwen3_vl_moe", "image-text-to-text", "conversational", "arxiv:2505.09388", "arxiv:2502.13923", "arxiv:2409.12191", "arxiv:2308.12966", "base_model:Qwen/Qwen3-VL-235B-A22B-Instruct", "base_model:quantized:Qwen/Qwen3-VL-235B-A22B-Instruct", "license:apache-2.0", "endpoints_compatible", "fp8", "deploy:azure", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_vl_moe_text", "total_params_b": 235.684292848, "active_params_b": 21.568433664, "layers": 94, "attention_heads": 64, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "FP8 E4M3 + BF16/F32 side tensors", "weight_bytes_per_param": 1.0081207753001782, "resident_weight_gb": 237.598232032, "active_weight_gb": 22.245946368, "kv_alloc_gb_per_1k_tokens": 0.192512, "kv_read_gb_per_1k_tokens": 0.192512, "notes": "Manually audited on 2026-07-07 from current HF API metadata, model card, pinned FP8 config, BF16 base profile comparison, official Transformers Qwen3 VL MoE implementation, safetensors index, and direct range-read shard headers. downloads=113778. Exact tensor payload bytes: 237.598232032 GB resident, 235.200795648 GB main text/logit resident, 2.397436384 GB resident-only visual/input embedding bytes, 8.048956416 GB fixed non-expert traffic, and 1.774623744 GB per routed expert. active_weight_gb is single-sequence ordinary text-decode traffic: fixed plus 8 routed experts. The audited profile supersedes these legacy adapter fields for expected-distinct expert batching." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Instruct-FP8" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Instruct-FP8/raw/7fbcd8c9e2ad011808ed8a57d64c462605c3e381/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-VL-235B-A22B-Instruct-FP8" }, { "label": "Base model profile", "url": "https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Instruct" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Instruct-FP8/resolve/7fbcd8c9e2ad011808ed8a57d64c462605c3e381/model.safetensors.index.json" } ] }, { "id": "ocicek/Qwen3.6-27B-NVFP4", "name": "Qwen3.6-27B-NVFP4 NVFP4", "short_name": "Qwen3.6 27B NVFP4", "license": "other", "hf_pipeline_tag": "text-generation", "hf_downloads": 117099, "tags": [ "vllm", "safetensors", "qwen3_5", "nvfp4", "compressed-tensors", "llm-compressor", "quantized", "qwen", "text-generation", "conversational", "base_model:Qwen/Qwen3.6-27B", "base_model:quantized:Qwen/Qwen3.6-27B", "license:other", "8-bit", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text", "total_params_b": 27.781427952, "active_params_b": 25.624600064, "layers": 64, "attention_heads": 24, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "NVFP4", "weight_bytes_per_param": 0.5, "resident_weight_gb": 20.558935392, "active_weight_gb": 16.245279616, "kv_alloc_gb_per_1k_tokens": 0.065536, "kv_read_gb_per_1k_tokens": 0.065536, "kv_alloc_fixed_gb": 0.154927104, "kv_read_fixed_gb": 0.154927104, "notes": "Manual correction on 2026-07-07 from current HF CLI/API metadata, model card, pinned compressed-tensors NVFP4 config and recipe, pinned base config comparison, and direct range-read safetensors headers for both indexed files. Live downloads were 84773 at audit time, but hf_downloads retains the original qualifying scrape count because the catalog schema is the over-100k catalog. Bound weights separate the 20.558935392 GB resident NVFP4/BF16/F8/F32 package from the 16.245279616 GB ordinary text-decode swept subset. Attention KV is charged as BF16 because the config has kv_cache_scheme null and the card's vLLM quick start does not request FP8 KV; fixed DeltaNet runtime state is included separately. The audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/ocicek/Qwen3.6-27B-NVFP4" }, { "label": "Model config", "url": "https://huggingface.co/ocicek/Qwen3.6-27B-NVFP4/raw/311613f3c18fe5dbc73a0fcaea7c721d258b1b4e/config.json" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/ocicek/Qwen3.6-27B-NVFP4" }, { "label": "Pinned NVFP4 recipe", "url": "https://huggingface.co/ocicek/Qwen3.6-27B-NVFP4/raw/311613f3c18fe5dbc73a0fcaea7c721d258b1b4e/recipe.yaml" }, { "label": "Pinned BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3.6-27B/raw/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/ocicek/Qwen3.6-27B-NVFP4/resolve/311613f3c18fe5dbc73a0fcaea7c721d258b1b4e/model.safetensors.index.json" } ] }, { "id": "ggml-org/gpt-oss-120b-GGUF", "name": "gpt-oss-120b GGUF MXFP4_MOE", "short_name": "gpt oss 120b", "license": "unknown", "hf_pipeline_tag": "unknown", "hf_downloads": 116699, "tags": [ "gguf", "base_model:openai/gpt-oss-120b", "base_model:quantized:openai/gpt-oss-120b", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "MoE transformer", "detail": "gpt_oss (GGUF MXFP4_MOE split)", "total_params_b": 116.829156672, "active_params_b": 5.1, "layers": 36, "attention_heads": 64, "kv_heads": 8, "head_dim": 64, "max_context_tokens": 131072, "routed_experts": 128, "routed_experts_per_token": 4, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "GGUF MXFP4_MOE/Q8_0", "weight_bytes_per_param": 0.5425644442676337, "resident_weight_gb": 63.387346464, "active_weight_gb": 3.594210048, "kv_alloc_gb_per_1k_tokens": 0.073728, "kv_read_gb_per_1k_tokens": 0.073728, "notes": "Manually corrected on 2026-07-07 from current Hugging Face API metadata, model card metadata, OpenAI base metadata, selected split linked-object HEAD checks, and direct GGUF header/tensor-index range reads of all three split files. The current live downloads count was 69411 at audit time, but the catalog keeps the original qualifying scrape count 116699 until the over-100k working set is regenerated. The authoritative profile targets the repo-level gpt-oss-120b-mxfp4 split package: 63.387346464 GB resident split size, 62.758994688 GB main resident tensor traffic, 0.628351776 GB resident-only input embedding plus GGUF overhead, and 3.594210048 GB single-session ordinary text-decode swept traffic." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/ggml-org/gpt-oss-120b-GGUF" }, { "label": "Hugging Face model card", "url": "https://huggingface.co/ggml-org/gpt-oss-120b-GGUF" }, { "label": "OpenAI base API metadata", "url": "https://huggingface.co/api/models/openai/gpt-oss-120b" }, { "label": "Split GGUF tree", "url": "https://huggingface.co/ggml-org/gpt-oss-120b-GGUF/tree/d932fcea62f83e088d8f076a2cd2d7eb02dfa682" }, { "label": "Selected split artifact", "url": "https://huggingface.co/ggml-org/gpt-oss-120b-GGUF/resolve/d932fcea62f83e088d8f076a2cd2d7eb02dfa682/gpt-oss-120b-mxfp4-00001-of-00003.gguf" } ] }, { "id": "unsloth/gemma-3-270m-it-GGUF", "name": "gemma-3-270m-it-GGUF F16", "short_name": "gemma 3 270m it", "license": "gemma", "hf_pipeline_tag": "text-generation", "hf_downloads": 105716, "tags": [ "transformers", "gguf", "gemma3", "unsloth", "gemma", "google", "text-generation", "arxiv:2503.19786", "arxiv:1905.07830", "arxiv:1905.10044", "arxiv:1911.11641", "arxiv:1705.03551", "arxiv:1911.01547", "arxiv:1907.10641", "arxiv:2311.07911", "arxiv:2311.12022", "arxiv:2411.04368", "arxiv:1904.09728", "arxiv:1903.00161", "arxiv:2009.03300", "arxiv:2304.06364", "arxiv:2103.03874", "arxiv:2110.14168", "arxiv:2108.07732", "arxiv:2107.03374", "arxiv:2403.07974", "arxiv:2305.03111", "arxiv:2405.04520", "arxiv:2210.03057", "arxiv:2106.03193", "arxiv:1910.11856", "arxiv:2502.12404", "arxiv:2502.21228", "arxiv:2404.16816", "arxiv:2104.12756", "arxiv:2311.16502", "arxiv:2203.10244", "arxiv:2404.12390", "arxiv:1810.12440", "arxiv:1908.02660", "arxiv:2310.02255", "arxiv:2312.11805", "base_model:google/gemma-3-270m-it", "base_model:quantized:google/gemma-3-270m-it", "license:gemma", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "gemma3 (F16 GGUF header)", "total_params_b": 0.268098176, "active_params_b": 0.268098176, "layers": 18, "attention_heads": 4, "kv_heads": 1, "head_dim": 256, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF F16/F32", "weight_bytes_per_param": 2, "resident_weight_gb": 0.542835488, "active_weight_gb": 0.536308224, "kv_alloc_gb_per_1k_tokens": 0.01093632, "kv_read_gb_per_1k_tokens": 0.01093632, "notes": "Audited 2026-07-07 from live HF API metadata, model card metadata, Google base API metadata, package side-file review, gated base-config access check, selected linked-object HEAD checks, and a range-read of the selected F16 GGUF header. API gguf.totalFileSize selects gemma-3-270m-it-F16.gguf. KV display fields show first-1K text-cache traffic; the audited profile supersedes these legacy adapter fields for production bounds and handles the 512-token local sliding-window cap." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/gemma-3-270m-it-GGUF" }, { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/gemma-3-270m-it-GGUF" }, { "label": "Base model API metadata", "url": "https://huggingface.co/api/models/google/gemma-3-270m-it" }, { "label": "Gated base config access check", "url": "https://huggingface.co/google/gemma-3-270m-it/raw/ac82b4e820549b854eebf28ce6dedaf9fdfa17b3/config.json" }, { "label": "Selected F16 GGUF artifact", "url": "https://huggingface.co/unsloth/gemma-3-270m-it-GGUF/resolve/c90975dbd40c0c7b275fefaae758c3415c906238/gemma-3-270m-it-F16.gguf" } ] }, { "id": "TheBloke/TinyLlama-1.1B-Chat-v0.3-AWQ", "name": "TinyLlama-1.1B-Chat-v0.3-AWQ 4-bit", "short_name": "TinyLlama 1.1B Chat v0.3", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 132292, "tags": [ "transformers", "safetensors", "llama", "text-generation", "en", "dataset:cerebras/SlimPajama-627B", "dataset:bigcode/starcoderdata", "dataset:OpenAssistant/oasst_top1_2023-08-25", "base_model:TinyLlama/TinyLlama-1.1B-Chat-v0.3", "base_model:quantized:TinyLlama/TinyLlama-1.1B-Chat-v0.3", "license:apache-2.0", "text-generation-inference", "4-bit", "awq", "deploy:azure", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "llama (LlamaForCausalLM AWQ)", "total_params_b": 1.100060672, "active_params_b": 1.034518528, "layers": 22, "attention_heads": 32, "kv_heads": 4, "head_dim": 64, "max_context_tokens": 2048, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "AWQ 4-bit group 128 GEMM", "weight_bytes_per_param": 0.6960693600725325, "resident_weight_gb": 0.765718528, "active_weight_gb": 0.63463424, "kv_alloc_gb_per_1k_tokens": 0.022528, "kv_read_gb_per_1k_tokens": 0.022528, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, model card, pinned AWQ config, quant_config, current TinyLlama base config comparison, and direct safetensors header inspection. Bound weights separate the exact 0.765718528 GB resident AWQ package from the 0.634634240 GB ordinary text-decode swept traffic; the audited profile supersedes the generated ideal 4-bit estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v0.3-AWQ" }, { "label": "Model config", "url": "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v0.3-AWQ/raw/d87af942d412211c2fe4a6cd6c4964cdd7e93e22/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v0.3" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/TheBloke/TinyLlama-1.1B-Chat-v0.3-AWQ" }, { "label": "Quant config", "url": "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v0.3-AWQ/raw/d87af942d412211c2fe4a6cd6c4964cdd7e93e22/quant_config.json" }, { "label": "Pinned base config", "url": "https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v0.3/raw/1ea93498c83d6d5aece90cf3b160aadf9f8875f8/config.json" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v0.3-AWQ/resolve/d87af942d412211c2fe4a6cd6c4964cdd7e93e22/model.safetensors" } ] }, { "id": "RedHatAI/Qwen2.5-VL-3B-Instruct-quantized.w4a16", "name": "Qwen2.5-VL-3B-Instruct-quantized.w4a16 4-bit", "short_name": "Qwen2.5 VL 3B Instruct quantized.w4a16", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 115971, "tags": [ "transformers", "safetensors", "qwen2_5_vl", "image-text-to-text", "vllm", "vision", "w4a16", "conversational", "en", "base_model:Qwen/Qwen2.5-VL-3B-Instruct", "base_model:quantized:Qwen/Qwen2.5-VL-3B-Instruct", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2_5_vl (compressed-tensors W4A16)", "total_params_b": 4.08746444, "active_params_b": 3.107615224, "layers": 36, "attention_heads": 16, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 128000, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "compressed-tensors W4A16 INT4/BF16", "weight_bytes_per_param": 0.9818149659547863, "resident_weight_gb": 4.01313376, "active_weight_gb": 2.053435328, "kv_alloc_gb_per_1k_tokens": 0.036864, "kv_read_gb_per_1k_tokens": 0.036864, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, model card, pinned compressed-tensors config, recipe.yaml, current Qwen2.5-VL 3B base config comparison, and direct safetensors header range reads. Bound weights separate the exact 4.013133760 GB resident tensor payload from the 2.053435328 GB ordinary text-decode swept traffic; the audited profile supersedes the generated ideal 4-bit estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/RedHatAI/Qwen2.5-VL-3B-Instruct-quantized.w4a16" }, { "label": "Model config", "url": "https://huggingface.co/RedHatAI/Qwen2.5-VL-3B-Instruct-quantized.w4a16/raw/a6bca3776bd02c77a80e66326c4441a651fdd2e7/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/RedHatAI/Qwen2.5-VL-3B-Instruct-quantized.w4a16" }, { "label": "Quantization recipe", "url": "https://huggingface.co/RedHatAI/Qwen2.5-VL-3B-Instruct-quantized.w4a16/raw/a6bca3776bd02c77a80e66326c4441a651fdd2e7/recipe.yaml" }, { "label": "Pinned base config", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct/raw/66285546d2b821cf421d4f5eb2576359d3770cd3/config.json" }, { "label": "Safetensors header audit", "url": "https://huggingface.co/RedHatAI/Qwen2.5-VL-3B-Instruct-quantized.w4a16/resolve/a6bca3776bd02c77a80e66326c4441a651fdd2e7/model.safetensors" } ] }, { "id": "nm-testing/Qwen3-Coder-30B-A3B-Instruct-W4A16-awq", "name": "Qwen3-Coder-30B-A3B-Instruct-W4A16-awq 4-bit", "short_name": "Qwen3 Coder 30B A3B Instruct W4A16", "license": "unknown", "hf_pipeline_tag": "unknown", "hf_downloads": 115566, "tags": [ "safetensors", "qwen3_moe", "base_model:Qwen/Qwen3-Coder-30B-A3B-Instruct", "base_model:quantized:Qwen/Qwen3-Coder-30B-A3B-Instruct", "compressed-tensors", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_moe (Qwen3MoeForCausalLM compressed-tensors W4A16)", "total_params_b": 30.532122624, "active_params_b": 3.3, "layers": 48, "attention_heads": 32, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "compressed-tensors W4A16 INT4/BF16", "weight_bytes_per_param": 0.5, "resident_weight_gb": 16.686185472, "active_weight_gb": 2.049360896, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "expert_param_b": 0.233574912, "shared_expert_param_b": 0, "fixed_param_b": 2.230122496, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, model card, pinned served compressed-tensors config, recipe.yaml, official BF16 base config comparison, safetensors index, linked-object metadata, and direct shard-header range reads. Bound weights use exact stored W4A16 tensor payload bytes: 16.686185472 GB resident package, 1.115061248 GB fixed ordinary traffic, 0.116787456 GB per routed expert index, and 0.622329856 GB resident-only input embedding. Current live downloads are 31877, but the older scraped qualifying count 115566 is retained for this audit queue." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nm-testing/Qwen3-Coder-30B-A3B-Instruct-W4A16-awq" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nm-testing/Qwen3-Coder-30B-A3B-Instruct-W4A16-awq" }, { "label": "Pinned served compressed-tensors config", "url": "https://huggingface.co/nm-testing/Qwen3-Coder-30B-A3B-Instruct-W4A16-awq/raw/9971cd6828ce3eefdcd9e9ca72dc4586ede07379/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct" }, { "label": "Quantization recipe", "url": "https://huggingface.co/nm-testing/Qwen3-Coder-30B-A3B-Instruct-W4A16-awq/raw/9971cd6828ce3eefdcd9e9ca72dc4586ede07379/recipe.yaml" }, { "label": "Safetensors index and shard-header audit", "url": "https://huggingface.co/nm-testing/Qwen3-Coder-30B-A3B-Instruct-W4A16-awq/resolve/9971cd6828ce3eefdcd9e9ca72dc4586ede07379/model.safetensors.index.json" } ] }, { "id": "unsloth/Qwen3.5-4B-MTP-GGUF", "name": "Qwen3.5-4B-MTP-GGUF BF16 GGUF", "short_name": "Qwen3.5 4B MTP", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 115139, "tags": [ "transformers", "gguf", "unsloth", "qwen", "qwen3_5", "image-text-to-text", "base_model:Qwen/Qwen3.5-4B", "base_model:quantized:Qwen/Qwen3.5-4B", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_5_text (selected BF16 GGUF with resident MTP draft block)", "total_params_b": 4.326350848, "active_params_b": 4.205751296, "layers": 32, "attention_heads": 16, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16/F32 MTP GGUF", "weight_bytes_per_param": 2.000450656007453, "resident_weight_gb": 8.665620544, "active_weight_gb": 8.413425664, "kv_alloc_gb_per_1k_tokens": 0.032768, "kv_read_gb_per_1k_tokens": 0.032768, "kv_alloc_fixed_gb": 0.051904512, "kv_read_fixed_gb": 0.051904512, "notes": "Manually audited on 2026-07-07 from HF API commit 86835bf9949e4d14d6860f7910b1340ad4f271a9, pinned Qwen base config, model card, linked GGUF file sizes, and a GGUF tensor-index range read. The authoritative bounds profile uses the selected Qwen3.5-4B-BF16.gguf artifact from the MTP repo: 8.665620544 GB resident, 8.413425664 GB swept ordinary text traffic, 0.241225728 GB resident-only MTP draft-block tensors, 0.010969152 GB resident-only GGUF metadata/header/file overhead, and Qwen3.5 hybrid KV plus fixed DeltaNet state. The tied token embedding is swept because no separate output.weight is stored. Current live downloads are 98149, but the older scraped qualifying count 115139 is retained for this audit queue." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/Qwen3.5-4B-MTP-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3.5-4B/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.5-4B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3.5-4B-MTP-GGUF" }, { "label": "Selected GGUF artifact", "url": "https://huggingface.co/unsloth/Qwen3.5-4B-MTP-GGUF/resolve/86835bf9949e4d14d6860f7910b1340ad4f271a9/Qwen3.5-4B-BF16.gguf" } ] }, { "id": "lmstudio-community/Qwen3-VL-4B-Instruct-MLX-4bit", "name": "Qwen3-VL-4B-Instruct-MLX-4bit BF16/U32", "short_name": "Qwen3 VL 4B Instruct MLX 4bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 110453, "tags": [ "mlx", "safetensors", "qwen3_vl", "image-text-to-text", "conversational", "base_model:Qwen/Qwen3-VL-4B-Instruct", "base_model:quantized:Qwen/Qwen3-VL-4B-Instruct", "license:apache-2.0", "4-bit", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_vl_text", "total_params_b": 4.563511808, "active_params_b": 4.148164096, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "MLX 4-bit affine + BF16 side tensors", "weight_bytes_per_param": 0.677902401956489, "resident_weight_gb": 3.093615616, "active_weight_gb": 2.262920192, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Audited on 2026-07-07 at commit 552af30c9952c44f1e1a27c7c5810ded58e892bc. Direct model.safetensors headers record 3.093615616 GB resident payload, with 2.262920192 GB language_model.model swept for ordinary tied-embedding text decode and 0.830695424 GB vision_tower resident-only. The checked-in index is stale and names two absent shards, so the actual single safetensors file is authoritative." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-4bit" }, { "label": "Model config", "url": "https://huggingface.co/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-4bit/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-4bit/revision/552af30c9952c44f1e1a27c7c5810ded58e892bc" }, { "label": "Audited safetensors artifact", "url": "https://huggingface.co/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-4bit/resolve/552af30c9952c44f1e1a27c7c5810ded58e892bc/model.safetensors" } ] }, { "id": "Qwen/Qwen2.5-14B-Instruct-GPTQ-Int4", "name": "Qwen2.5-14B-Instruct-GPTQ-Int4 4-bit", "short_name": "Qwen2.5 14B Instruct GPTQ Int4", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 111341, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "chat", "conversational", "en", "arxiv:2309.00071", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-14B-Instruct", "base_model:quantized:Qwen/Qwen2.5-14B-Instruct", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "4-bit", "gptq", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 14.770033664, "active_params_b": 13.991465984, "layers": 48, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GPTQ Int4 + F16 tensors", "weight_bytes_per_param": 0.676273433306103, "resident_weight_gb": 9.988581376, "active_weight_gb": 8.431446016, "kv_alloc_gb_per_1k_tokens": 0.196608, "kv_read_gb_per_1k_tokens": 0.196608, "notes": "Audited on 2026-07-07 at commit 08fdbde5eea15cd2aae884125fbfd604dec19818. Direct safetensors shard headers record 9.988581376 GB resident payload and 8.431446016 GB ordinary swept text-decode traffic. The GPTQ package stores qweight, qzeros, g_idx, F16 scales/biases, and unquantized F16 embedding/head/norm tensors." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-14B-Instruct-GPTQ-Int4" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen2.5-14B-Instruct-GPTQ-Int4/raw/08fdbde5eea15cd2aae884125fbfd604dec19818/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-14B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-14B-Instruct-GPTQ-Int4" }, { "label": "Audited safetensors index", "url": "https://huggingface.co/Qwen/Qwen2.5-14B-Instruct-GPTQ-Int4/resolve/08fdbde5eea15cd2aae884125fbfd604dec19818/model.safetensors.index.json" } ] }, { "id": "bartowski/Qwen_Qwen3.6-35B-A3B-GGUF", "name": "Qwen_Qwen3.6-35B-A3B-GGUF GGUF", "short_name": "Qwen Qwen3.6 35B A3B", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 120082, "tags": [ "gguf", "image-text-to-text", "base_model:Qwen/Qwen3.6-35B-A3B", "base_model:quantized:Qwen/Qwen3.6-35B-A3B", "license:apache-2.0", "endpoints_compatible", "region:us", "imatrix", "conversational" ], "architecture": { "type": "MoE transformer", "detail": "qwen35moe / qwen3_5_moe_text", "total_params_b": 35.505251456, "active_params_b": 3, "layers": 40, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "GGUF IQ1_M", "weight_bytes_per_param": 0.265359324315047, "resident_weight_gb": 9.421649536, "active_weight_gb": 1.428179456, "kv_alloc_gb_per_1k_tokens": 0.02048, "kv_read_gb_per_1k_tokens": 0.02048, "kv_alloc_fixed_gb": 0.06488064, "kv_read_fixed_gb": 0.06488064, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, model card, pinned Qwen base config, selected linked-object HEAD checks, and direct GGUF header range read. The selected API artifact is Qwen_Qwen3.6-35B-A3B-IQ1_M.gguf at 9.421649536 GB resident; the model card explicitly marks IQ1_M as not recommended and Q4_K_M as the default recommended size. Ordinary single-session swept traffic is 1.428179456 GB before KV/state reads, with blk.40 MTP tensors treated as resident-only for ordinary non-speculative text decode." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/bartowski/Qwen_Qwen3.6-35B-A3B-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/bartowski/Qwen_Qwen3.6-35B-A3B-GGUF" }, { "label": "Selected GGUF header audit", "url": "https://huggingface.co/bartowski/Qwen_Qwen3.6-35B-A3B-GGUF/resolve/5c2410d71524f4f72b023ce8daf7a80528226d5f/Qwen_Qwen3.6-35B-A3B-IQ1_M.gguf" } ] }, { "id": "unsloth/GLM-4.7-Flash-GGUF", "name": "GLM-4.7-Flash-GGUF GGUF", "short_name": "GLM 4.7 Flash", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 112163, "tags": [ "transformers", "gguf", "unsloth", "text-generation", "en", "zh", "arxiv:2508.06471", "base_model:zai-org/GLM-4.7-Flash", "base_model:quantized:zai-org/GLM-4.7-Flash", "license:mit", "endpoints_compatible", "deploy:azure", "region:us", "imatrix", "conversational" ], "architecture": { "type": "Latent MLA MoE transformer", "detail": "deepseek2 GGUF / glm4_moe_lite", "total_params_b": 29.94339392, "active_params_b": 3, "layers": 47, "attention_heads": 20, "kv_heads": 1, "head_dim": 576, "max_context_tokens": 202752, "routed_experts": 64, "routed_experts_per_token": 4, "shared_experts_per_token": 1, "kv_lora_rank": 512, "qk_rope_head_dim": 64, "qk_nope_head_dim": 192, "v_head_dim": 256 }, "adapter": { "kind": "moe", "weight_precision": "GGUF IQ4_NL", "weight_bytes_per_param": 0.573249440923763, "resident_weight_gb": 17.165033824, "active_weight_gb": 2.325907968, "kv_alloc_gb_per_1k_tokens": 0.054144, "kv_read_gb_per_1k_tokens": 0.054144, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, pinned Unsloth card, pinned Z.ai base config, linked-object HEAD checks, direct GGUF header range read, and current llama.cpp DeepSeek2/cache source review. The catalog retains the older 112163-download snapshot that qualified this row; live API downloads were 94041 during audit. Bound weights use the exact 17.165033824 GB API-selected IQ4_NL linked artifact. Active single-session traffic charges current llama.cpp ordinary decode: 1.349159424 GB fixed tensors plus 4 routed experts at 0.244187136 GB each. The selected main file has no blk.47, MTP/NextN, DSA indexer, or multimodal sidecar tensors." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/GLM-4.7-Flash-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/zai-org/GLM-4.7-Flash/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/zai-org/GLM-4.7-Flash" }, { "label": "Pinned Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/GLM-4.7-Flash-GGUF" }, { "label": "Selected IQ4_NL GGUF header audit", "url": "https://huggingface.co/unsloth/GLM-4.7-Flash-GGUF/resolve/0d32489ecb9db6d2a4fc93bd27ef01519f95474d/GLM-4.7-Flash-IQ4_NL.gguf" }, { "label": "llama.cpp DeepSeek2 source review", "url": "https://github.com/ggml-org/llama.cpp/blob/cb295bf59663cd3577389315636772f4060bd1f5/src/models/deepseek2.cpp" } ] }, { "id": "meta-llama/Llama-2-13b-chat-hf", "name": "Llama-2-13b-chat-hf F32/F16", "short_name": "Llama 2 13b chat", "license": "llama2", "hf_pipeline_tag": "text-generation", "hf_downloads": 115324, "tags": [ "transformers", "pytorch", "safetensors", "llama", "text-generation", "facebook", "meta", "llama-2", "conversational", "en", "arxiv:2307.09288", "license:llama2", "text-generation-inference", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "gated config unavailable", "total_params_b": 13.01586688, "active_params_b": 13.01586688, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "F32/F16", "weight_bytes_per_param": 2.0000003933660393, "resident_weight_gb": 26.03173888, "active_weight_gb": 26.03173888, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manual correction on 2026-07-07 from current Hugging Face API metadata and gated access checks. The API reports F16 plus tiny F32 safetensors counts, but raw config and model.safetensors.index.json are gated from the configured osolmaz CLI token. Production bounds use the fail-closed unsupported profile until config and tensor evidence are directly accessible." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/meta-llama/Llama-2-13b-chat-hf" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-2-13b-chat-hf" } ] }, { "id": "lmstudio-community/Qwen3-VL-4B-Instruct-MLX-8bit", "name": "Qwen3-VL-4B-Instruct-MLX-8bit BF16/U32", "short_name": "Qwen3 VL 4B Instruct MLX 8bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 106978, "tags": [ "mlx", "safetensors", "qwen3_vl", "image-text-to-text", "conversational", "base_model:Qwen/Qwen3-VL-4B-Instruct", "base_model:quantized:Qwen/Qwen3-VL-4B-Instruct", "license:apache-2.0", "8-bit", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_vl_text", "total_params_b": 4.563511808, "active_params_b": 4.148164096, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "MLX 8-bit affine + BF16 side tensors", "weight_bytes_per_param": 1.1186016013043258, "resident_weight_gb": 5.104751616, "active_weight_gb": 4.274056192, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Audited on 2026-07-07 at commit e73e3fbb7dae9f23283989a81a05d327a3958f3f. Direct model.safetensors headers record 5.104751616 GB resident payload, with 4.274056192 GB language_model.model swept for ordinary tied-embedding text decode and 0.830695424 GB vision_tower resident-only. The checked-in index is stale and names two absent shards, so the actual single safetensors file is authoritative." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-8bit" }, { "label": "Model config", "url": "https://huggingface.co/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-8bit/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-8bit/revision/e73e3fbb7dae9f23283989a81a05d327a3958f3f" }, { "label": "Audited safetensors artifact", "url": "https://huggingface.co/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-8bit/resolve/e73e3fbb7dae9f23283989a81a05d327a3958f3f/model.safetensors" } ] }, { "id": "lmstudio-community/Qwen3-VL-4B-Instruct-MLX-6bit", "name": "Qwen3-VL-4B-Instruct-MLX-6bit BF16/U32", "short_name": "Qwen3 VL 4B Instruct MLX 6bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 106558, "tags": [ "mlx", "safetensors", "qwen3_vl", "image-text-to-text", "conversational", "base_model:Qwen/Qwen3-VL-4B-Instruct", "base_model:quantized:Qwen/Qwen3-VL-4B-Instruct", "license:apache-2.0", "6-bit", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_vl_text", "total_params_b": 4.563511808, "active_params_b": 4.148164096, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "MLX 6-bit affine + BF16 side tensors", "weight_bytes_per_param": 0.8982520016304075, "resident_weight_gb": 4.099183616, "active_weight_gb": 3.268488192, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Audited on 2026-07-07 at commit 31afb1d6a62ebb3ed2ff43643c5ce22368d6d85a. Direct model.safetensors headers record 4.099183616 GB resident payload, with 3.268488192 GB language_model.model swept for ordinary tied-embedding text decode and 0.830695424 GB vision_tower resident-only. The checked-in index is stale and names two absent shards, so the actual single safetensors file is authoritative." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-6bit" }, { "label": "Model config", "url": "https://huggingface.co/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-6bit/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-6bit/revision/31afb1d6a62ebb3ed2ff43643c5ce22368d6d85a" }, { "label": "Audited safetensors artifact", "url": "https://huggingface.co/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-6bit/resolve/31afb1d6a62ebb3ed2ff43643c5ce22368d6d85a/model.safetensors" } ] }, { "id": "lmstudio-community/Qwen3-VL-4B-Instruct-MLX-5bit", "name": "Qwen3-VL-4B-Instruct-MLX-5bit BF16/U32", "short_name": "Qwen3 VL 4B Instruct MLX 5bit", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 106372, "tags": [ "mlx", "safetensors", "qwen3_vl", "image-text-to-text", "conversational", "base_model:Qwen/Qwen3-VL-4B-Instruct", "base_model:quantized:Qwen/Qwen3-VL-4B-Instruct", "license:apache-2.0", "5-bit", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3_vl_text", "total_params_b": 4.563511808, "active_params_b": 4.148164096, "layers": 36, "attention_heads": 32, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "MLX 5-bit affine + BF16 side tensors", "weight_bytes_per_param": 0.7880772017934482, "resident_weight_gb": 3.596399616, "active_weight_gb": 2.765704192, "kv_alloc_gb_per_1k_tokens": 0.147456, "kv_read_gb_per_1k_tokens": 0.147456, "notes": "Audited on 2026-07-07 at commit 2a6fe7e4f6cd3ee80b3dbb5b4deeb4183f6707b0. Direct model.safetensors headers record 3.596399616 GB resident payload, with 2.765704192 GB language_model.model swept for ordinary tied-embedding text decode and 0.830695424 GB vision_tower resident-only. The checked-in index is stale and names two absent shards, so the actual single safetensors file is authoritative." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-5bit" }, { "label": "Model config", "url": "https://huggingface.co/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-5bit/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-5bit/revision/2a6fe7e4f6cd3ee80b3dbb5b4deeb4183f6707b0" }, { "label": "Audited safetensors artifact", "url": "https://huggingface.co/lmstudio-community/Qwen3-VL-4B-Instruct-MLX-5bit/resolve/2a6fe7e4f6cd3ee80b3dbb5b4deeb4183f6707b0/model.safetensors" } ] }, { "id": "cyankiwi/Qwen3-Coder-30B-A3B-Instruct-AWQ-4bit", "name": "Qwen3-Coder-30B-A3B-Instruct-AWQ-4bit 4-bit", "short_name": "Qwen3 Coder 30B A3B Instruct AWQ 4bit", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 245223, "tags": [ "transformers", "safetensors", "qwen3_moe", "text-generation", "conversational", "arxiv:2505.09388", "base_model:Qwen/Qwen3-Coder-30B-A3B-Instruct", "base_model:quantized:Qwen/Qwen3-Coder-30B-A3B-Instruct", "license:apache-2.0", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_moe (Qwen3MoeForCausalLM)", "total_params_b": 30.532122624, "active_params_b": 3.3, "layers": 48, "attention_heads": 32, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "compressed-tensors AWQ INT4/BF16 group-size 32", "weight_bytes_per_param": 0.5, "resident_weight_gb": 18.087607296, "active_weight_gb": 2.17676288, "kv_alloc_gb_per_1k_tokens": 0.098304, "kv_read_gb_per_1k_tokens": 0.098304, "expert_param_b": 0.254808576, "shared_expert_param_b": 0, "fixed_param_b": 2.315057152, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, model card, pinned served compressed-tensors AWQ config, recipe.yaml, official BF16 base config comparison, safetensors index, linked-object metadata, and direct shard-header range reads. Bound weights use exact stored AWQ tensor payload bytes: 18.087607296 GB resident package, 1.157528576 GB fixed ordinary traffic, 0.127404288 GB per routed expert index, and 0.622329856 GB resident-only input embedding. The repo metadata identifies the official Qwen3-Coder base, but served config differences mean the audited profile is authoritative for this package. Current live downloads are 245223." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/cyankiwi/Qwen3-Coder-30B-A3B-Instruct-AWQ-4bit" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3-Coder-30B-A3B-Instruct-AWQ-4bit" }, { "label": "Pinned served compressed-tensors AWQ config", "url": "https://huggingface.co/cyankiwi/Qwen3-Coder-30B-A3B-Instruct-AWQ-4bit/raw/2831070b7b8c7aa6b7012333c6c4a2bd257f6cdf/config.json" }, { "label": "Official BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct/raw/b2cff646eb4bb1d68355c01b18ae02e7cf42d120/config.json" }, { "label": "Quantization recipe", "url": "https://huggingface.co/cyankiwi/Qwen3-Coder-30B-A3B-Instruct-AWQ-4bit/raw/2831070b7b8c7aa6b7012333c6c4a2bd257f6cdf/recipe.yaml" }, { "label": "Safetensors index and shard-header audit", "url": "https://huggingface.co/cyankiwi/Qwen3-Coder-30B-A3B-Instruct-AWQ-4bit/resolve/2831070b7b8c7aa6b7012333c6c4a2bd257f6cdf/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen2.5-32B-Instruct-GPTQ-Int8", "name": "Qwen2.5-32B-Instruct-GPTQ-Int8 8-bit", "short_name": "Qwen2.5 32B Instruct GPTQ Int8", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 118191, "tags": [ "transformers", "safetensors", "qwen2", "text-generation", "chat", "conversational", "en", "arxiv:2309.00071", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-32B-Instruct", "base_model:quantized:Qwen/Qwen2.5-32B-Instruct", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "8-bit", "gptq", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 (Qwen2ForCausalLM)", "total_params_b": 32.763876352, "active_params_b": 31.985308672, "layers": 64, "attention_heads": 40, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 32768, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GPTQ 8-bit + F16 side tensors", "weight_bytes_per_param": 1.070339007974535, "resident_weight_gb": 35.068454912, "active_weight_gb": 33.511319552, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Manual correction on 2026-07-07 from current Hugging Face API metadata, served GPTQ config, model card, current base config comparison, safetensors index, linked-object HEAD checks, and direct safetensors header range reads. Bound weights separate the 35.068454912 GB resident GPTQ package from the 33.511319552 GB ordinary text-decode swept subset; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct-GPTQ-Int8" }, { "label": "Pinned served GPTQ config", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct-GPTQ-Int8/raw/eddc13f573fd3648cc8a4741fdf1b70e8d6fc5c1/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen2.5-32B-Instruct-GPTQ-Int8" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct-GPTQ-Int8/resolve/eddc13f573fd3648cc8a4741fdf1b70e8d6fc5c1/model.safetensors.index.json" } ] }, { "id": "unsloth/gemma-3-12b-it-GGUF", "name": "gemma-3-12b-it-GGUF BF16", "short_name": "gemma 3 12b it", "license": "gemma", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 108980, "tags": [ "transformers", "gguf", "gemma3", "image-text-to-text", "unsloth", "gemma", "google", "en", "arxiv:1905.07830", "arxiv:1905.10044", "arxiv:1911.11641", "arxiv:1904.09728", "arxiv:1705.03551", "arxiv:1911.01547", "arxiv:1907.10641", "arxiv:1903.00161", "arxiv:2009.03300", "arxiv:2304.06364", "arxiv:2103.03874", "arxiv:2110.14168", "arxiv:2311.12022", "arxiv:2108.07732", "arxiv:2107.03374", "arxiv:2210.03057", "arxiv:2106.03193", "arxiv:1910.11856", "arxiv:2502.12404", "arxiv:2502.21228", "arxiv:2404.16816", "arxiv:2104.12756", "arxiv:2311.16502", "arxiv:2203.10244", "arxiv:2404.12390", "arxiv:1810.12440", "arxiv:1908.02660", "arxiv:2312.11805", "base_model:google/gemma-3-12b-it", "base_model:quantized:google/gemma-3-12b-it", "license:gemma", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "gemma3_text (GGUF BF16 selected artifact)", "total_params_b": 11.766034176, "active_params_b": 11.766034176, "layers": 48, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF BF16/F32", "weight_bytes_per_param": 2, "resident_weight_gb": 23.54015152, "active_weight_gb": 23.533599744, "kv_alloc_gb_per_1k_tokens": 0.065536, "kv_read_gb_per_1k_tokens": 0.065536, "kv_alloc_fixed_gb": 0.33554432, "kv_read_fixed_gb": 0.33554432, "notes": "Audited 2026-07-07 from live HF API metadata, package Gemma 3 config, model card, gated-base access check, exact linked-object HEAD checks, Google Gemma 3 local/global attention documentation, and a range-read of the selected BF16 GGUF header. The catalog retains the older qualifying scrape count 108980 while the live API now reports 97138 downloads. API gguf.totalFileSize selects gemma-3-12b-it-BF16.gguf, so this row intentionally models the BF16 artifact rather than lower-bit GGUF siblings advertised by tags. Exact selected file residency is 23.540151520 GB and ordinary swept tensor traffic is 23.533599744 GB. KV uses Gemma 3 layered FP16 K/V over eight full-context layers plus 40 sliding-window layers capped at 1024 tokens." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/gemma-3-12b-it-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/gemma-3-12b-it-GGUF" }, { "label": "Package Gemma 3 config", "url": "https://huggingface.co/unsloth/gemma-3-12b-it-GGUF/raw/d15e4c7dc21dc55d56bf8549db57a71ad8a2a35d/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-3-12b-it" }, { "label": "Selected BF16 GGUF artifact", "url": "https://huggingface.co/unsloth/gemma-3-12b-it-GGUF/resolve/d15e4c7dc21dc55d56bf8549db57a71ad8a2a35d/gemma-3-12b-it-BF16.gguf" } ] }, { "id": "cyankiwi/Nex-N2-mini-AWQ-INT4", "name": "Nex-N2-mini-AWQ-INT4 4-bit", "short_name": "Nex N2 mini AWQ INT4", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 108667, "tags": [ "transformers", "safetensors", "qwen3_5_moe", "image-text-to-text", "text-generation", "conversational", "base_model:nex-agi/Nex-N2-mini", "base_model:quantized:nex-agi/Nex-N2-mini", "license:apache-2.0", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen35moe / qwen3_5_moe_text", "total_params_b": 37.120509296, "active_params_b": 3, "layers": 40, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "compressed-tensors AWQ INT4", "weight_bytes_per_param": 0.6576702791798893, "resident_weight_gb": 24.413055712, "active_weight_gb": 4.461568256, "kv_alloc_gb_per_1k_tokens": 0.02048, "kv_read_gb_per_1k_tokens": 0.02048, "kv_alloc_fixed_gb": 0.06488064, "kv_read_fixed_gb": 0.06488064, "notes": "Manually audited on 2026-07-07 from live Hugging Face API metadata, model card, pinned served compressed-tensors config, base config comparison, linked-object HEAD checks, and direct safetensors header range reads. Exact resident payload is 24.413055712 GB; ordinary single-session swept traffic is 4.461568256 GB before KV/state reads. The audited profile supersedes the generated ideal 4-bit estimate and charges the Qwen3.5 hybrid KV plus fixed DeltaNet state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/cyankiwi/Nex-N2-mini-AWQ-INT4" }, { "label": "Model config", "url": "https://huggingface.co/cyankiwi/Nex-N2-mini-AWQ-INT4/raw/00ff7ea4cd56acf5ed4c0d8144fb6cb68765b446/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/nex-agi/Nex-N2-mini" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Nex-N2-mini-AWQ-INT4" }, { "label": "Safetensors index", "url": "https://huggingface.co/cyankiwi/Nex-N2-mini-AWQ-INT4/resolve/00ff7ea4cd56acf5ed4c0d8144fb6cb68765b446/model.safetensors.index.json" } ] }, { "id": "cyankiwi/Qwen3-Coder-Next-AWQ-8bit", "name": "Qwen3-Coder-Next-AWQ-8bit 8-bit", "short_name": "Qwen3 Coder Next AWQ 8bit", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 108320, "tags": [ "transformers", "safetensors", "qwen3_next", "text-generation", "conversational", "base_model:Qwen/Qwen3-Coder-Next", "base_model:quantized:Qwen/Qwen3-Coder-Next", "license:apache-2.0", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_next (Qwen3NextForCausalLM)", "total_params_b": 79.674391296, "active_params_b": 3.56376448, "layers": 48, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "compressed-tensors AWQ INT8", "weight_bytes_per_param": 1.090342677326, "resident_weight_gb": 86.87238912, "active_weight_gb": 5.7119744, "kv_alloc_gb_per_1k_tokens": 0.024576, "kv_read_gb_per_1k_tokens": 0.024576, "kv_alloc_fixed_gb": 0.077856768, "kv_read_fixed_gb": 0.077856768, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, pinned compressed-tensors AWQ config and recipe, official BF16 base config comparison, safetensors index, direct shard-header range reads, and the Transformers qwen3_next implementation. Bound weights use the exact 86.872389120 GB resident package and 5.711974400 GB single-session ordinary text-decode traffic; the audited profile supersedes the generated ideal 8-bit estimate, fixes shared expert count, and adds the fixed DeltaNet runtime state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/cyankiwi/Qwen3-Coder-Next-AWQ-8bit" }, { "label": "Pinned Hugging Face API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3-Coder-Next-AWQ-8bit" }, { "label": "Pinned model config", "url": "https://huggingface.co/cyankiwi/Qwen3-Coder-Next-AWQ-8bit/raw/a5a598b8531a5ef032db7242cf7f5b49ab2eada6/config.json" }, { "label": "Pinned compression recipe", "url": "https://huggingface.co/cyankiwi/Qwen3-Coder-Next-AWQ-8bit/raw/a5a598b8531a5ef032db7242cf7f5b49ab2eada6/recipe.yaml" }, { "label": "Pinned BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3-Coder-Next/raw/a7fbcb5c0e12d62a448eaa0e260346bf5dcc0feb/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/cyankiwi/Qwen3-Coder-Next-AWQ-8bit/resolve/a5a598b8531a5ef032db7242cf7f5b49ab2eada6/model.safetensors.index.json" }, { "label": "Transformers Qwen3-Next implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/qwen3_next/modeling_qwen3_next.py" } ] }, { "id": "Qwen/Qwen2.5-7B-Instruct-GGUF", "name": "Qwen2.5-7B-Instruct-GGUF GGUF Q2_K", "short_name": "Qwen2.5 7B Instruct", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 107559, "tags": [ "gguf", "chat", "text-generation", "en", "arxiv:2407.10671", "base_model:Qwen/Qwen2.5-7B-Instruct", "base_model:quantized:Qwen/Qwen2.5-7B-Instruct", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen2 GGUF (Qwen2ForCausalLM)", "total_params_b": 7.615616512, "active_params_b": 7.070619136, "layers": 28, "attention_heads": 28, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q2_K", "weight_bytes_per_param": 0.3960204660053135, "resident_weight_gb": 3.01594, "active_weight_gb": 2.831159296, "kv_alloc_gb_per_1k_tokens": 0.057344, "kv_read_gb_per_1k_tokens": 0.057344, "notes": "Audited on 2026-07-07 from current Hugging Face API metadata, pinned model card, audited BF16 Instruct profile, linked-object HEAD checks, and direct selected Q2_K GGUF tensor-index range read. The catalog snapshot downloads count is retained for threshold validation; the live API count during audit was 99520. The API gguf.totalFileSize points at qwen2.5-7b-instruct-q2_k.gguf, so Q2_K is the selected artifact. Bound weights use exact selected-file bytes: 3.015940000 GB resident, 2.831159296 GB ordinary text swept traffic, and 0.184780704 GB resident-only token embedding plus GGUF overhead." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct" } ] }, { "id": "Jackrong/Qwopus3.6-27B-Coder-Compat-MTP-GGUF", "name": "Qwopus3.6-27B-Coder-Compat-MTP-GGUF GGUF Q4_K_M", "short_name": "Qwopus3.6 27B Coder Compat MTP", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 322097, "tags": [ "transformers", "gguf", "llama.cpp", "image-text-to-text", "vision", "multimodal", "text-generation-inference", "unsloth", "conversational", "qwen3_6", "reasoning", "chain-of-thought", "agent", "tool-use", "function-calling", "coder", "mtp", "speculative-decoding", "compatibility", "en", "zh", "es", "ru", "ja", "dataset:Jackrong/Claude-opus-4.6-TraceInversion-9000x", "dataset:Jackrong/Claude-opus-4.7-TraceInversion-5000x", "dataset:lambda/hermes-agent-reasoning-traces", "base_model:Jackrong/Qwopus3.6-27B-Coder", "base_model:quantized:Jackrong/Qwopus3.6-27B-Coder", "license:apache-2.0", "endpoints_compatible", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen35 GGUF (Qwen3_5ForConditionalGeneration + MTP)", "total_params_b": 27.320697856, "active_params_b": 25.624600064, "layers": 64, "attention_heads": 24, "kv_heads": 4, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "GGUF Q4_K_M", "weight_bytes_per_param": 0.6153104214469447, "resident_weight_gb": 16.810710112, "active_weight_gb": 15.821244416, "kv_alloc_gb_per_1k_tokens": 0.065536, "kv_read_gb_per_1k_tokens": 0.065536, "kv_alloc_fixed_gb": 0.154927104, "kv_read_fixed_gb": 0.154927104, "notes": "Manually audited on 2026-07-07 from current Hugging Face API metadata, pinned model card, pinned Jackrong Coder config, selected linked-object HEAD checks, direct selected Q4_K_M GGUF tensor-index range read, and Qwen3.5 runtime review. The API GGUF block points at mmproj-F32 with architecture clip, so the audited profile uses the Compat-specific card-validated Q4_K_M main artifact. Bound weights use exact selected-file bytes: 16.810710112 GB resident, 15.821244416 GB ordinary text swept traffic, and 0.989465696 GB resident-only token embedding plus MTP/header overhead. KV traffic uses Qwen3.6 hybrid full-context KV plus fixed DeltaNet recurrent state." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Jackrong/Qwopus3.6-27B-Coder-Compat-MTP-GGUF/raw/2b74b86b39dda302b0cb165145d2b0a48c31e833/README.md" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Jackrong/Qwopus3.6-27B-Coder-Compat-MTP-GGUF" }, { "label": "Pinned base model config", "url": "https://huggingface.co/Jackrong/Qwopus3.6-27B-Coder/raw/1eee80aef82e96900a3d3aa88c302e0dbf0e5e49/config.json" }, { "label": "Selected Q4_K_M GGUF tensor-index artifact", "url": "https://huggingface.co/Jackrong/Qwopus3.6-27B-Coder-Compat-MTP-GGUF/resolve/2b74b86b39dda302b0cb165145d2b0a48c31e833/Qwopus3.6-27B-Coder-Compat-MTP-Q4_K_M.gguf" } ] }, { "id": "RecViking/Mistral-Medium-3.5-128B-NVFP4", "name": "Mistral-Medium-3.5-128B-NVFP4 NVFP4", "short_name": "Mistral Medium 3.5 128B NVFP4", "license": "other", "hf_pipeline_tag": "unknown", "hf_downloads": 106484, "tags": [ "safetensors", "mistral3", "vLLM", "compressed-tensors", "nvfp4", "quantized", "mistral", "multimodal", "en", "fr", "de", "es", "pt", "it", "ja", "ko", "ru", "zh", "ar", "fa", "id", "ms", "ne", "pl", "ro", "sr", "sv", "tr", "uk", "vi", "hi", "bn", "base_model:mistralai/Mistral-Medium-3.5-128B", "base_model:quantized:mistralai/Mistral-Medium-3.5-128B", "license:other", "8-bit", "region:us" ], "architecture": { "type": "Dense multimodal transformer", "detail": "mistral3 / ministral3 text", "total_params_b": 127.704210176, "active_params_b": 123.415375872, "layers": 88, "attention_heads": 96, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "NVFP4", "weight_bytes_per_param": 0.5, "resident_weight_gb": 80.3172048, "active_weight_gb": 71.739536192, "kv_alloc_gb_per_1k_tokens": 0.090112, "kv_read_gb_per_1k_tokens": 0.090112, "notes": "Manually audited on 2026-07-08 from current Hugging Face API metadata, pinned model card, pinned served config, pinned recipe, pinned base config comparison, and direct two-shard safetensors header range reads. Live downloads were 18825 at audit time, but hf_downloads retains the original qualifying scrape count 106484 until the over-100k working set is regenerated. Bound weights use exact tensor payload bytes: 80.317204800 GB resident package, 71.739536192 GB ordinary text-decode swept subset, and 8.577668608 GB resident-only embedding, vision, and projector tensors. KV display fields use the documented vLLM --kv-cache-dtype turboquant_4bit_nc launch; the audited profile supersedes these legacy adapter fields for production bounds." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/RecViking/Mistral-Medium-3.5-128B-NVFP4" }, { "label": "Model config", "url": "https://huggingface.co/RecViking/Mistral-Medium-3.5-128B-NVFP4/raw/dc2c3c2962bf46c3a5335d19b1f4b4725011da32/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/mistralai/Mistral-Medium-3.5-128B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/RecViking/Mistral-Medium-3.5-128B-NVFP4" }, { "label": "Pinned NVFP4 recipe", "url": "https://huggingface.co/RecViking/Mistral-Medium-3.5-128B-NVFP4/raw/dc2c3c2962bf46c3a5335d19b1f4b4725011da32/recipe.yaml" }, { "label": "Pinned BF16 base config", "url": "https://huggingface.co/mistralai/Mistral-Medium-3.5-128B/raw/ed0c85631cf20929c29bc36d7cb37519ed0cdd28/config.json" }, { "label": "Safetensors index and shard-header audit", "url": "https://huggingface.co/RecViking/Mistral-Medium-3.5-128B-NVFP4/resolve/dc2c3c2962bf46c3a5335d19b1f4b4725011da32/model.safetensors.index.json" } ] }, { "id": "unsloth/Step-3.7-Flash-GGUF", "name": "Step-3.7-Flash-GGUF GGUF UD-Q4_K_S", "short_name": "Step 3.7 Flash", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 105722, "tags": [ "transformers", "gguf", "vision-language", "unsloth - multimodal - moe", "image-text-to-text", "en", "base_model:stepfun-ai/Step-3.7-Flash", "base_model:quantized:stepfun-ai/Step-3.7-Flash", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "MoE transformer", "detail": "step35 GGUF / Step3p7 text MoE", "total_params_b": 196.956130432, "active_params_b": 11, "layers": 45, "attention_heads": 64, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 288, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "GGUF UD-Q4_K_S/Q8_0/Q6_K/F32", "weight_bytes_per_param": 0.579637669554111, "resident_weight_gb": 114.163192448, "active_weight_gb": 9.552003584, "kv_alloc_gb_per_1k_tokens": 0.049152, "kv_read_gb_per_1k_tokens": 0.049152, "kv_alloc_fixed_gb": 0.069206016, "kv_read_fixed_gb": 0.069206016, "expert_param_b": 0.641071378756054, "shared_expert_param_b": 0, "fixed_param_b": 11.35069539055519, "notes": "Manually audited on 2026-07-08 from current Hugging Face API metadata, pinned Unsloth model card, pinned StepFun base config, linked-object HEAD checks, and direct selected GGUF split-header range reads. Catalog downloads retain the original qualifying scrape count 105722; live API downloads at audit time were 44668. The profile targets the card-recommended local-serving UD-Q4_K_S split rather than the API BF16 aggregate. Bound weights use exact linked/tensor bytes: 114.163192448 GB resident split package, 6.579290624 GB fixed ordinary traffic, 0.371589120 GB per routed expert index across 288 experts, and 0.566235264 GB resident-only token embedding plus GGUF overhead. KV traffic uses FP16 full/sliding cache: 12 full-context layers plus 33 sliding-window layers capped at 512 tokens." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/Step-3.7-Flash-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/Step-3.7-Flash-GGUF" }, { "label": "Pinned StepFun base config", "url": "https://huggingface.co/stepfun-ai/Step-3.7-Flash/raw/5f6244077ac62e04eec3f320501ff8c2b293373a/config.json" }, { "label": "Selected UD-Q4_K_S split audit", "url": "https://huggingface.co/unsloth/Step-3.7-Flash-GGUF/tree/8dd62fe8a98f3c8994160d0c521542560b558f37/UD-Q4_K_S" } ] }, { "id": "lmstudio-community/GLM-4.6V-Flash-MLX-4bit", "name": "GLM-4.6V-Flash-MLX-4bit MLX 4-bit", "short_name": "GLM 4.6V Flash MLX 4bit", "license": "mit", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 111225, "tags": [ "transformers", "safetensors", "glm4v", "image-text-to-text", "mlx", "conversational", "zh", "en", "base_model:zai-org/GLM-4.6V-Flash", "base_model:quantized:zai-org/GLM-4.6V-Flash", "license:mit", "endpoints_compatible", "4-bit", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "glm4v_text with resident GLM4V vision tower", "total_params_b": 10.292777472, "active_params_b": 8.779522048, "layers": 40, "attention_heads": 32, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "MLX 4-bit affine + BF16 side tensors", "weight_bytes_per_param": 0.6872651013046209, "resident_weight_gb": 7.073866752, "active_weight_gb": 4.93969408, "kv_alloc_gb_per_1k_tokens": 0.04096, "kv_read_gb_per_1k_tokens": 0.04096, "notes": "Manually audited on 2026-07-08 from current Hugging Face API metadata, pinned LM Studio card/config, current Z.ai base config, stale index review, direct two-shard safetensors headers, and pinned mlx-vlm GLM4V/cache source review. The checked-in index names four stale shards that 404; the actual linked files are model-00001-of-00002.safetensors and model-00002-of-00002.safetensors. Bound weights use exact stored MLX bytes: 7.073866752 GB resident payload, 4.939694080 GB ordinary text-decode swept traffic, and 2.134172672 GB resident-only token embedding plus vision tower." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/GLM-4.6V-Flash-MLX-4bit" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/GLM-4.6V-Flash-MLX-4bit" }, { "label": "Pinned MLX config", "url": "https://huggingface.co/lmstudio-community/GLM-4.6V-Flash-MLX-4bit/raw/f3626bba332c159be847ee1a4e68a5fcbde915ae/config.json" }, { "label": "Current BF16 base config", "url": "https://huggingface.co/zai-org/GLM-4.6V-Flash/raw/411bb4d77144a3f03accbf4b780f5acb8b7cde4e/config.json" }, { "label": "Actual safetensors shard audit", "url": "https://huggingface.co/lmstudio-community/GLM-4.6V-Flash-MLX-4bit/tree/f3626bba332c159be847ee1a4e68a5fcbde915ae" } ] }, { "id": "bartowski/google_gemma-4-26B-A4B-it-GGUF", "name": "Gemma 4 26B A4B IT GGUF IQ2_M", "short_name": "Gemma 4 26B A4B IT GGUF", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 106081, "tags": [ "gguf", "image-text-to-text", "base_model:google/gemma-4-26B-A4B-it", "base_model:quantized:google/gemma-4-26B-A4B-it", "license:apache-2.0", "endpoints_compatible", "region:us", "imatrix", "conversational" ], "architecture": { "type": "MoE transformer", "detail": "gemma4_text", "total_params_b": 25.233142046, "active_params_b": 3.8, "layers": 30, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "GGUF IQ2_M mixed", "weight_bytes_per_param": 0.424070085, "resident_weight_gb": 10.700620704, "active_weight_gb": 2.04619712, "kv_alloc_gb_per_1k_tokens": 0.20992, "kv_read_gb_per_1k_tokens": 0.20992, "notes": "Manual correction on 2026-07-08 from current HF API metadata, selected GGUF linked-object HEAD checks, existing Google Gemma 4 profile/config geometry, and a direct GGUF tensor-index range read. The authoritative bounds profile uses the API-selected google_gemma-4-26B-A4B-it-IQ2_M.gguf artifact: 10.700620704 GB resident, 1.470290368 GB fixed traffic including tied token_embd/output projection, 0.071988344 GB per routed expert, 8 routed experts per token, 128 routed experts, 1 shared expert charged in fixed traffic, and layered Gemma 4 hybrid KV. Legacy per-1k KV fields show the short-context unsaturated coefficient; long-context calculation must use the profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/bartowski/google_gemma-4-26B-A4B-it-GGUF" }, { "label": "Base model config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/bartowski/google_gemma-4-26B-A4B-it-GGUF" }, { "label": "Selected IQ2_M GGUF artifact", "url": "https://huggingface.co/bartowski/google_gemma-4-26B-A4B-it-GGUF/resolve/fabed3e586120477355eea23b92644540a79ce2f/google_gemma-4-26B-A4B-it-IQ2_M.gguf" } ] }, { "id": "Jiunsong/supergemma4-26b-uncensored-gguf-v2", "name": "SuperGemma4 26B Uncensored Fast v2 GGUF Q4_K_M", "short_name": "SuperGemma4 26B Fast v2", "license": "gemma", "hf_pipeline_tag": "text-generation", "hf_downloads": 104781, "tags": [ "gguf", "gemma4", "uncensored", "fast", "llama.cpp", "apple-silicon", "conversational", "korean", "coding", "tool-use", "text-generation", "en", "ko", "base_model:google/gemma-4-26B-A4B-it", "base_model:quantized:google/gemma-4-26B-A4B-it", "license:gemma", "endpoints_compatible", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "gemma4_text", "total_params_b": 25.233142046, "active_params_b": 3.8, "layers": 30, "attention_heads": 16, "kv_heads": 8, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "GGUF Q4_K_M mixed", "weight_bytes_per_param": 0.665633127, "resident_weight_gb": 16.796015232, "active_weight_gb": 2.595649408, "kv_alloc_gb_per_1k_tokens": 0.20992, "kv_read_gb_per_1k_tokens": 0.20992, "notes": "Manual correction on 2026-07-08 from current HF API metadata, selected GGUF linked-object HEAD checks, model card metadata, existing Google Gemma 4 profile/config geometry, and a direct GGUF tensor-index range read. The catalog retains the original qualifying scrape count 104781 because live API downloads were 86717 at audit time. The authoritative bounds profile uses the single selected supergemma4-26b-uncensored-fast-v2-Q4_K_M.gguf artifact: 16.796015232 GB resident, 1.650013120 GB fixed traffic including tied token_embd/output projection, 0.118204536 GB per routed expert, 8 routed experts per token, 128 routed experts, 1 shared expert charged in fixed traffic, and layered Gemma 4 hybrid KV. Legacy per-1k KV fields show the short-context unsaturated coefficient; long-context calculation must use the profile." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Jiunsong/supergemma4-26b-uncensored-gguf-v2" }, { "label": "Base model config", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it/raw/main/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-26B-A4B-it" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Jiunsong/supergemma4-26b-uncensored-gguf-v2" }, { "label": "Selected Q4_K_M GGUF artifact", "url": "https://huggingface.co/Jiunsong/supergemma4-26b-uncensored-gguf-v2/resolve/3ea8c452a2b136875c0c8b529612bed39c81e27a/supergemma4-26b-uncensored-fast-v2-Q4_K_M.gguf" } ] }, { "id": "nvidia/GLM-5.2-NVFP4", "name": "GLM-5.2-NVFP4 NVFP4", "short_name": "GLM 5.2 NVFP4", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 449881, "tags": [ "Model Optimizer", "safetensors", "glm_moe_dsa", "nvidia", "ModelOpt", "GLM-5", "quantized", "4-bit precision", "FP4", "fp4", "text-generation", "conversational", "base_model:zai-org/GLM-5.2", "base_model:quantized:zai-org/GLM-5.2", "license:mit", "8-bit", "modelopt", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "glm_moe_dsa (GlmMoeDsaForCausalLM)", "total_params_b": 753.32994048, "active_params_b": 40, "layers": 78, "attention_heads": 64, "kv_heads": 64, "head_dim": 256, "max_context_tokens": 1048576, "routed_experts": 256, "routed_experts_per_token": 8, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "ModelOpt NVFP4 + BF16 fixed tensors", "weight_bytes_per_param": 0.6169876465760088, "resident_weight_gb": 464.795267072, "active_weight_gb": 48.039663168, "kv_alloc_gb_per_1k_tokens": 2.558592, "kv_read_gb_per_1k_tokens": 2.558592, "expert_param_b": 2.581132067777651, "fixed_param_b": 57.21257234872586, "shared_expert_param_b": 0, "notes": "Manual correction on 2026-07-08 from current Hugging Face API metadata, pinned model card, target config, hf_quant_config, base GLM-5.2 config comparison, safetensors index, and direct range-read safetensors shard headers. Active traffic is the single-session expected-distinct MoE text-decode weight traffic. KV display fields charge FP8 expanded K/V for 78 layers plus FP8 GLM-5.2 IndexShare DSA indexer key cache for 21 full indexer layers. The audited profile supersedes the generated 376.6649 GB resident estimate and all-layers BF16-ish KV estimate. Corrected on 2026-07-09: the routed expert group size was re-derived as 1.592526600 GB after the earlier split double-counted shared-expert tensors inside the routed sum." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/GLM-5.2-NVFP4" }, { "label": "Model config", "url": "https://huggingface.co/nvidia/GLM-5.2-NVFP4/raw/aec724e8c7b8ee9db3b48c01c320f63f9cdaf8aa/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/zai-org/GLM-5.2" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/GLM-5.2-NVFP4" }, { "label": "Safetensors index", "url": "https://huggingface.co/nvidia/GLM-5.2-NVFP4/resolve/aec724e8c7b8ee9db3b48c01c320f63f9cdaf8aa/model.safetensors.index.json" } ] }, { "id": "meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8", "name": "Llama-4-Maverick-17B-128E-Instruct-FP8 FP8", "short_name": "Llama 4 Maverick 17B 128E Instruct FP8", "license": "other", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 108194, "tags": [ "transformers", "safetensors", "llama4", "image-text-to-text", "facebook", "meta", "pytorch", "llama", "conversational", "ar", "de", "en", "es", "fr", "hi", "id", "it", "pt", "th", "tl", "vi", "arxiv:2204.05149", "base_model:meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8", "base_model:quantized:meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8", "license:other", "eval-results", "text-generation-inference", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "Unknown gated multimodal MoE", "detail": "gated compressed-tensors FP8 config unavailable", "total_params_b": 401.649841664, "active_params_b": 401.649841664, "layers": null, "attention_heads": null, "kv_heads": null, "head_dim": null, "max_context_tokens": null, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "FP8 + BF16 side tensors", "weight_bytes_per_param": 1.037601869731681, "resident_weight_gb": 416.752626688, "active_weight_gb": 416.752626688, "kv_alloc_gb_per_1k_tokens": 0, "kv_read_gb_per_1k_tokens": 0, "notes": "Manual fail-closed correction on 2026-07-08 from current Hugging Face API metadata and direct raw-file/HF CLI gated-access checks. The API records 401.649841664B safetensors parameters at commit 94125d2bd83076b21eed33119525e29eaf3894f4: 15.102785024B BF16 and 386.547056640B F8_E4M3, or 416.752626688 GB if charged by stored dtype. The shallow API config exposes Llama4ForConditionalGeneration/model_type llama4 and compressed-tensors 8-bit Linear quantization. Raw config, README, generation_config.json, recipe.yaml, model.safetensors.index.json, and hf download config.json all require accepted-gate access in this environment. Production bounds use the unsupported profile and must not infer KV geometry, MoE routing, context, or swept traffic from this row." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8" }, { "label": "Gated config access check", "url": "https://huggingface.co/meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8/raw/94125d2bd83076b21eed33119525e29eaf3894f4/config.json" }, { "label": "Gated tensor index access check", "url": "https://huggingface.co/meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8/raw/94125d2bd83076b21eed33119525e29eaf3894f4/model.safetensors.index.json" } ] }, { "id": "lmstudio-community/GLM-4.6V-Flash-MLX-8bit", "name": "GLM-4.6V-Flash-MLX-8bit MLX 8-bit", "short_name": "GLM 4.6V Flash MLX 8bit", "license": "mit", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 109777, "tags": [ "transformers", "safetensors", "glm4v", "image-text-to-text", "mlx", "conversational", "zh", "en", "base_model:zai-org/GLM-4.6V-Flash", "base_model:quantized:zai-org/GLM-4.6V-Flash", "license:mit", "endpoints_compatible", "8-bit", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "glm4v_text with resident GLM4V vision tower", "total_params_b": 10.292777472, "active_params_b": 8.779522048, "layers": 40, "attention_heads": 32, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "MLX 8-bit affine + BF16 side tensors", "weight_bytes_per_param": 1.1438685443291006, "resident_weight_gb": 11.773584384, "active_weight_gb": 9.329033216, "kv_alloc_gb_per_1k_tokens": 0.04096, "kv_read_gb_per_1k_tokens": 0.04096, "notes": "Manually audited on 2026-07-08 from current Hugging Face API metadata, pinned LM Studio card/config, current Z.ai base config, stale index review, direct three-shard safetensors headers, and pinned mlx-vlm GLM4V/cache source review. The checked-in index names four stale shards that 404; the actual linked files are model-00001-of-00003.safetensors, model-00002-of-00003.safetensors, and model-00003-of-00003.safetensors. Bound weights use exact stored MLX bytes: 11.773584384 GB resident payload, 9.329033216 GB ordinary text-decode swept traffic, and 2.444551168 GB resident-only token embedding plus vision tower." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/GLM-4.6V-Flash-MLX-8bit" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/GLM-4.6V-Flash-MLX-8bit" }, { "label": "Pinned MLX config", "url": "https://huggingface.co/lmstudio-community/GLM-4.6V-Flash-MLX-8bit/raw/77ab1c1729e2011cc22bdda37b5bc43cc3e22b1d/config.json" }, { "label": "Current BF16 base config", "url": "https://huggingface.co/zai-org/GLM-4.6V-Flash/raw/411bb4d77144a3f03accbf4b780f5acb8b7cde4e/config.json" }, { "label": "Actual safetensors shard audit", "url": "https://huggingface.co/lmstudio-community/GLM-4.6V-Flash-MLX-8bit/tree/77ab1c1729e2011cc22bdda37b5bc43cc3e22b1d" } ] }, { "id": "cyankiwi/Qwen3-Next-80B-A3B-Thinking-AWQ-4bit", "name": "Qwen3-Next-80B-A3B-Thinking-AWQ-4bit 4-bit", "short_name": "Qwen3 Next 80B A3B Thinking AWQ 4bit", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 105276, "tags": [ "transformers", "safetensors", "qwen3_next", "text-generation", "conversational", "arxiv:2309.00071", "arxiv:2505.09388", "arxiv:2501.15383", "base_model:Qwen/Qwen3-Next-80B-A3B-Thinking", "base_model:quantized:Qwen/Qwen3-Next-80B-A3B-Thinking", "license:apache-2.0", "endpoints_compatible", "compressed-tensors", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_next (Qwen3NextForCausalLM, compressed-tensors AWQ)", "total_params_b": 83.791264, "active_params_b": 3, "layers": 48, "attention_heads": 16, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 262144, "routed_experts": 512, "routed_experts_per_token": 10, "shared_experts_per_token": 1 }, "adapter": { "kind": "moe", "weight_precision": "compressed-tensors AWQ Int4 packed I32 + BF16/I64 side tensors", "weight_bytes_per_param": 0.5872139018215551, "resident_weight_gb": 49.203395072, "active_weight_gb": 4.95699968, "kv_alloc_gb_per_1k_tokens": 0.024576, "kv_read_gb_per_1k_tokens": 0.024576, "kv_alloc_fixed_gb": 0.077856768, "kv_read_fixed_gb": 0.077856768, "expert_param_b": 0.14464398703185163, "fixed_param_b": 6.995117226036387, "shared_expert_param_b": 0, "notes": "Manually corrected on 2026-07-08 from the scraped catalog row, current Hugging Face API metadata, pinned compressed-tensors config, current BF16 base config comparison, recipe.yaml, safetensors index, and direct range-read safetensors shard headers. Legacy fallback values mirror the audited profile: exact stored resident bytes, fixed always-on hybrid/shared-expert traffic, expected distinct routed experts, BF16 attention KV, and fixed BF16-convolution/F32-SSM DeltaNet recurrent state. The authoritative profile supersedes the generated 41.8957 GB resident estimate, full-context-only KV estimate, and missing shared-expert count." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/cyankiwi/Qwen3-Next-80B-A3B-Thinking-AWQ-4bit" }, { "label": "Model config", "url": "https://huggingface.co/cyankiwi/Qwen3-Next-80B-A3B-Thinking-AWQ-4bit/raw/1c7d46f509dc053c42befba435aecb39ce6a7921/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Thinking" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/cyankiwi/Qwen3-Next-80B-A3B-Thinking-AWQ-4bit" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/cyankiwi/Qwen3-Next-80B-A3B-Thinking-AWQ-4bit/resolve/1c7d46f509dc053c42befba435aecb39ce6a7921/model.safetensors.index.json" }, { "label": "Transformers Qwen3-Next implementation", "url": "https://github.com/huggingface/transformers/blob/9f909e0f7027285d83bb6addc88155f1b80244ab/src/transformers/models/qwen3_next/modeling_qwen3_next.py" } ] }, { "id": "lmstudio-community/GLM-4.6V-Flash-MLX-6bit", "name": "GLM-4.6V-Flash-MLX-6bit MLX 6-bit", "short_name": "GLM 4.6V Flash MLX 6bit", "license": "mit", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 108813, "tags": [ "transformers", "safetensors", "glm4v", "image-text-to-text", "mlx", "conversational", "zh", "en", "base_model:zai-org/GLM-4.6V-Flash", "base_model:quantized:zai-org/GLM-4.6V-Flash", "license:mit", "endpoints_compatible", "6-bit", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "glm4v_text with resident GLM4V vision tower", "total_params_b": 10.292777472, "active_params_b": 8.779522048, "layers": 40, "attention_heads": 32, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "MLX 6-bit affine + BF16 side tensors", "weight_bytes_per_param": 0.9155668228168607, "resident_weight_gb": 9.423725568, "active_weight_gb": 7.134363648, "kv_alloc_gb_per_1k_tokens": 0.04096, "kv_read_gb_per_1k_tokens": 0.04096, "notes": "Manually audited on 2026-07-08 from current Hugging Face API metadata, pinned LM Studio card/config, current Z.ai base config, stale index review, direct two-shard safetensors headers, and pinned mlx-vlm GLM4V/cache source review. The checked-in index names four stale shards that 404; the actual linked files are model-00001-of-00002.safetensors and model-00002-of-00002.safetensors. Bound weights use exact stored MLX bytes: 9.423725568 GB resident payload, 7.134363648 GB ordinary text-decode swept traffic including lm_head side tensors, and 2.289361920 GB resident-only token embedding plus vision tower." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/lmstudio-community/GLM-4.6V-Flash-MLX-6bit" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/lmstudio-community/GLM-4.6V-Flash-MLX-6bit" }, { "label": "Pinned MLX config", "url": "https://huggingface.co/lmstudio-community/GLM-4.6V-Flash-MLX-6bit/raw/485d12f4f975842648828f8904939650f130a8d5/config.json" }, { "label": "Current BF16 base config", "url": "https://huggingface.co/zai-org/GLM-4.6V-Flash/raw/411bb4d77144a3f03accbf4b780f5acb8b7cde4e/config.json" }, { "label": "Actual safetensors shard audit", "url": "https://huggingface.co/lmstudio-community/GLM-4.6V-Flash-MLX-6bit/tree/485d12f4f975842648828f8904939650f130a8d5" } ] }, { "id": "nvidia/Qwen3-VL-235B-A22B-Instruct-NVFP4-MLPerf-Inference-Closed-V6.0", "name": "Qwen3-VL-235B-A22B-Instruct-NVFP4-MLPerf-Inference-Closed-V6.0 NVFP4", "short_name": "Qwen3 VL 235B A22B Instruct NVFP4 MLPerf Inferen", "license": "unknown", "hf_pipeline_tag": "unknown", "hf_downloads": 106913, "tags": [ "safetensors", "qwen3_vl_moe", "8-bit", "compressed-tensors", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "qwen3_vl_moe_text (compressed-tensors NVFP4 MLPerf package)", "total_params_b": 133.383079904, "active_params_b": 22, "layers": 94, "attention_heads": 64, "kv_heads": 4, "head_dim": 128, "max_context_tokens": 262144, "routed_experts": 128, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "NVFP4 compressed-tensors + BF16/F8/F32 side tensors", "weight_bytes_per_param": 1.0140298766631184, "resident_weight_gb": 135.254428064, "active_weight_gb": 13.098856, "kv_alloc_gb_per_1k_tokens": 0.192512, "kv_read_gb_per_1k_tokens": 0.192512, "notes": "Manually audited on 2026-07-08 from current Hugging Face API metadata, pinned served NVFP4 config, recipe.yaml, audited Qwen3-VL BF16 base profile/config comparison, Transformers Qwen3-VL MoE source review, safetensors index, and direct range-read safetensors shard headers across all 28 shards. Bound weights use exact stored compressed-tensors bytes: 135.254428064 GB resident payload, 5.114980288 GB fixed ordinary text/logit traffic, 0.997984464 GB per routed expert index, 13.098856000 GB single-session active traffic, and 2.397436384 GB resident-only input embedding plus vision tower." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/nvidia/Qwen3-VL-235B-A22B-Instruct-NVFP4-MLPerf-Inference-Closed-V6.0" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/nvidia/Qwen3-VL-235B-A22B-Instruct-NVFP4-MLPerf-Inference-Closed-V6.0" }, { "label": "Pinned NVFP4 config", "url": "https://huggingface.co/nvidia/Qwen3-VL-235B-A22B-Instruct-NVFP4-MLPerf-Inference-Closed-V6.0/raw/4875dbf337c70bf4cf44b712375bab7bd77eab3e/config.json" }, { "label": "Pinned NVFP4 recipe", "url": "https://huggingface.co/nvidia/Qwen3-VL-235B-A22B-Instruct-NVFP4-MLPerf-Inference-Closed-V6.0/raw/4875dbf337c70bf4cf44b712375bab7bd77eab3e/recipe.yaml" }, { "label": "Audited BF16 base config", "url": "https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Instruct/raw/710c13861be6c466e66de3f484069440b8f31389/config.json" }, { "label": "Safetensors index and header audit", "url": "https://huggingface.co/nvidia/Qwen3-VL-235B-A22B-Instruct-NVFP4-MLPerf-Inference-Closed-V6.0/resolve/4875dbf337c70bf4cf44b712375bab7bd77eab3e/model.safetensors.index.json" } ] }, { "id": "llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic-GGUF", "name": "gemma-4-E4B-it-ultra-uncensored-heretic-GGUF BF16 GGUF", "short_name": "gemma 4 E4B it ultra uncensored heretic", "license": "apache-2.0", "hf_pipeline_tag": "any-to-any", "hf_downloads": 102868, "tags": [ "transformers", "gguf", "gemma4", "image-text-to-text", "heretic", "uncensored", "decensored", "abliterated", "ara", "any-to-any", "base_model:llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic", "base_model:quantized:llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "gemma4_text with resident per-layer embedding", "total_params_b": 7.51806929, "active_params_b": 4.699497002, "layers": 42, "attention_heads": 8, "kv_heads": 2, "head_dim": 256, "max_context_tokens": 131072, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16 GGUF", "weight_bytes_per_param": 2.0022520541384323, "resident_weight_gb": 15.053094752, "active_weight_gb": 9.400126784, "kv_alloc_gb_per_1k_tokens": 0.086016, "kv_read_gb_per_1k_tokens": 0.086016, "notes": "Manually audited on 2026-07-08 from current Hugging Face API metadata, pinned model card, accessible source derivative config, existing Google Gemma 4 E4B profile/config convention, linked-object HEAD checks, and a direct GGUF header/tensor-index range read of the selected BF16 artifact. The repo now has 92138 live downloads, below the original over-100k scrape threshold; the catalog keeps the original 102868 scrape count because catalog rows currently require at least 100000 downloads. Bound weights use the selected BF16 main GGUF file size 15.053094752 GB for residency and 9.400126784 GB of ordinary swept tensor traffic after excluding the resident per-layer token embedding lookup table. The separate 0.991552000 GB mmproj sidecar is not included unless explicitly loaded." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic-GGUF" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic-GGUF" }, { "label": "Pinned source config", "url": "https://huggingface.co/llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic/raw/5964fe4c7339c5974e879baba8982a09616f68ca/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/google/gemma-4-E4B-it" }, { "label": "Selected BF16 GGUF artifact", "url": "https://huggingface.co/llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic-GGUF/resolve/1465f37b7dbd15e91241ae78ffebecb9f25e15de/gemma-4-E4B-it-ultra-uncensored-heretic-BF16.gguf" } ] }, { "id": "RedHatAI/Qwen2.5-VL-3B-Instruct-quantized.w8a8", "name": "Qwen2.5-VL-3B-Instruct-quantized.w8a8 8-bit", "short_name": "Qwen2.5 VL 3B Instruct quantized.w8a8", "license": "apache-2.0", "hf_pipeline_tag": "image-text-to-text", "hf_downloads": 121846, "tags": [ "transformers", "safetensors", "qwen2_5_vl", "image-text-to-text", "vllm", "vision", "w8a8", "conversational", "en", "base_model:Qwen/Qwen2.5-VL-3B-Instruct", "base_model:quantized:Qwen/Qwen2.5-VL-3B-Instruct", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "8-bit", "compressed-tensors", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen2_5_vl (compressed-tensors W8A8)", "total_params_b": 4.066820096, "active_params_b": 3.08697088, "layers": 36, "attention_heads": 16, "kv_heads": 2, "head_dim": 128, "max_context_tokens": 128000, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "compressed-tensors W8A8 INT8/BF16", "weight_bytes_per_param": 1.317763748947502, "resident_weight_gb": 5.359108096, "active_weight_gb": 3.399409664, "kv_alloc_gb_per_1k_tokens": 0.036864, "kv_read_gb_per_1k_tokens": 0.036864, "notes": "Manually audited on 2026-07-08 from current Hugging Face API metadata, model card, pinned compressed-tensors config, recipe.yaml, current Qwen2.5-VL 3B base config comparison, and direct safetensors header range reads across both shards. Bound weights separate the exact 5.359108096 GB resident tensor payload from the 3.399409664 GB ordinary text-decode swept traffic; the audited profile supersedes the generated ideal 8-bit estimate." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/RedHatAI/Qwen2.5-VL-3B-Instruct-quantized.w8a8" }, { "label": "Model config", "url": "https://huggingface.co/RedHatAI/Qwen2.5-VL-3B-Instruct-quantized.w8a8/raw/fcd4bf97086d0a782ead353e4ea700db0628b568/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/RedHatAI/Qwen2.5-VL-3B-Instruct-quantized.w8a8" }, { "label": "Quantization recipe", "url": "https://huggingface.co/RedHatAI/Qwen2.5-VL-3B-Instruct-quantized.w8a8/raw/fcd4bf97086d0a782ead353e4ea700db0628b568/recipe.yaml" }, { "label": "Pinned base config", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct/raw/66285546d2b821cf421d4f5eb2576359d3770cd3/config.json" }, { "label": "Safetensors index/header audit", "url": "https://huggingface.co/RedHatAI/Qwen2.5-VL-3B-Instruct-quantized.w8a8/resolve/fcd4bf97086d0a782ead353e4ea700db0628b568/model.safetensors.index.json" } ] }, { "id": "Qwen/Qwen3-32B-FP8", "name": "Qwen3-32B-FP8 FP8", "short_name": "Qwen3 32B FP8", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 101235, "tags": [ "transformers", "safetensors", "qwen3", "text-generation", "conversational", "arxiv:2309.00071", "arxiv:2505.09388", "base_model:Qwen/Qwen3-32B", "base_model:quantized:Qwen/Qwen3-32B", "license:apache-2.0", "text-generation-inference", "endpoints_compatible", "fp8", "region:us" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 (Qwen3ForCausalLM)", "total_params_b": 32.764027904, "active_params_b": 31.986115584, "layers": 64, "attention_heads": 64, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "FP8", "weight_bytes_per_param": 1.0475645469649273, "resident_weight_gb": 34.322434048, "active_weight_gb": 32.766609408, "kv_alloc_gb_per_1k_tokens": 0.262144, "kv_read_gb_per_1k_tokens": 0.262144, "notes": "Manually audited on 2026-07-08 from current Hugging Face API metadata, pinned served FP8 config, model card, audited BF16 base config/profile comparison, safetensors index, and direct shard header range reads. Live API downloads are now 99892, below the original scrape threshold; the catalog keeps the original 101235 scrape count because catalog rows currently require at least 100000 downloads. Bound weights use exact direct tensor payload bytes: 34.322434048 GB resident and 32.766609408 GB ordinary text-decode swept traffic after excluding only model.embed_tokens.weight." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/Qwen/Qwen3-32B-FP8" }, { "label": "Model config", "url": "https://huggingface.co/Qwen/Qwen3-32B-FP8/raw/aa55da1ecc13d006e8b8e4f54579b1ea8c3db2df/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-32B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/Qwen/Qwen3-32B-FP8" }, { "label": "Pinned base config", "url": "https://huggingface.co/Qwen/Qwen3-32B/raw/9216db5781bf21249d130ec9da846c4624c16137/config.json" }, { "label": "Safetensors index/header audit", "url": "https://huggingface.co/Qwen/Qwen3-32B-FP8/raw/aa55da1ecc13d006e8b8e4f54579b1ea8c3db2df/model.safetensors.index.json" } ] }, { "id": "unsloth/Qwen3-0.6B-GGUF", "name": "Qwen3-0.6B-GGUF GGUF", "short_name": "Qwen3 0.6B", "license": "apache-2.0", "hf_pipeline_tag": "text-generation", "hf_downloads": 100231, "tags": [ "transformers", "gguf", "qwen3", "text-generation", "qwen", "unsloth", "en", "base_model:Qwen/Qwen3-0.6B", "base_model:quantized:Qwen/Qwen3-0.6B", "license:apache-2.0", "endpoints_compatible", "region:us", "conversational" ], "architecture": { "type": "Dense transformer", "detail": "qwen3 GGUF (tied BF16)", "total_params_b": 0.59604992, "active_params_b": 0.59604992, "layers": 28, "attention_heads": 16, "kv_heads": 8, "head_dim": 128, "max_context_tokens": 40960, "routed_experts": null, "routed_experts_per_token": null, "shared_experts_per_token": null }, "adapter": { "kind": "dense", "weight_precision": "BF16 GGUF", "weight_bytes_per_param": 2.010205534462617, "resident_weight_gb": 1.198182848, "active_weight_gb": 1.192230912, "kv_alloc_gb_per_1k_tokens": 0.114688, "kv_read_gb_per_1k_tokens": 0.114688, "notes": "Manually audited on 2026-07-08 from current Hugging Face API metadata, pinned model card/config, linked-object HEAD checks, audited Qwen3 0.6B base profile comparison, and a direct GGUF header/tensor-index range read of the selected BF16 artifact. Bound weights use exact direct file/tensor bytes: 1.198182848 GB resident and 1.192230912 GB ordinary text-decode swept traffic. The selected GGUF omits output.weight, so token_embd.weight is swept as tied output-projection traffic." }, "sources": [ { "label": "Hugging Face model card", "url": "https://huggingface.co/unsloth/Qwen3-0.6B-GGUF" }, { "label": "Model config", "url": "https://huggingface.co/unsloth/Qwen3-0.6B-GGUF/raw/50968a4468ef4233ed78cd7c3de230dd1d61a56b/config.json" }, { "label": "Base model card", "url": "https://huggingface.co/Qwen/Qwen3-0.6B" }, { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/unsloth/Qwen3-0.6B-GGUF" }, { "label": "Selected BF16 GGUF artifact", "url": "https://huggingface.co/unsloth/Qwen3-0.6B-GGUF/resolve/50968a4468ef4233ed78cd7c3de230dd1d61a56b/Qwen3-0.6B-BF16.gguf" } ] }, { "id": "XiaomiMiMo/MiMo-V2.5-Pro", "name": "MiMo-V2.5-Pro FP8", "short_name": "MiMo V2.5 Pro", "license": "mit", "hf_pipeline_tag": "text-generation", "hf_downloads": 101692, "tags": [ "safetensors", "mimo_v2", "text-generation", "agent", "long-context", "code", "conversational", "custom_code", "en", "zh", "license:mit", "eval-results", "fp8", "region:us" ], "architecture": { "type": "MoE transformer", "detail": "mimo_v2 hybrid full/SWA MoE", "total_params_b": 1023.244718976, "active_params_b": 40.959508608, "layers": 70, "attention_heads": 128, "kv_heads": 8, "head_dim": 192, "max_context_tokens": 1048576, "routed_experts": 384, "routed_experts_per_token": 8, "shared_experts_per_token": 0 }, "adapter": { "kind": "moe", "weight_precision": "FP8", "weight_bytes_per_param": 1.009894817085528, "resident_weight_gb": 1033.369538304, "active_weight_gb": 49.438742016, "kv_alloc_gb_per_1k_tokens": 0.051593216, "kv_read_gb_per_1k_tokens": 0.0524288, "notes": "Audited from pinned Hugging Face metadata, config, custom code, and direct safetensors shard-header range reads on 2026-07-08; downloads=101692. Resident and active weights use exact stored FP8/BF16/F32 bytes; KV is hybrid full-context plus 128-token sliding window BF16." }, "sources": [ { "label": "Hugging Face API metadata", "url": "https://huggingface.co/api/models/XiaomiMiMo/MiMo-V2.5-Pro" }, { "label": "Pinned model card", "url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro/raw/f5953f1b993e90367f3d87fed59d099f2c527763/README.md" }, { "label": "Pinned model config", "url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro/raw/f5953f1b993e90367f3d87fed59d099f2c527763/config.json" }, { "label": "Pinned custom modeling code", "url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro/raw/f5953f1b993e90367f3d87fed59d099f2c527763/modeling_mimo_v2.py" }, { "label": "Pinned safetensors index", "url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro/raw/f5953f1b993e90367f3d87fed59d099f2c527763/model.safetensors.index.json" } ] } ] };