Qwen3.5-4B-mentria

Qwen3.5-4B in Q4 safetensors format for mentria-engine — a custom WebGPU runtime that runs the model entirely in your browser. Includes a Q4 vision tower for image input and a hot-swappable LoRA fine-tuned for motivational quotes.

This is the top tier of the mentria model ladder (0.8B2B → 4B) — served only to high-capability devices (≥2 GiB GPU buffer support, ≥8 GB device memory); other devices fall back to a smaller tier.

Try it live:

Files

Path Size Purpose
qwen3.5-4b-q4-tied-00001-of-00002.safetensors ~1.9 GB Language model, shard 1 of 2 — Q4_0 weights (MSE-optimal per-block scale) with tied embedding/lm_head. Sharded under 2 GB so each shard fits a JS ArrayBuffer. Loaded with allowTiedEmbed.
qwen3.5-4b-q4-tied-00002-of-00002.safetensors ~750 MB Language model, shard 2 of 2
qwen3.5-4b-vl-q4.safetensors ~210 MB Vision tower — Q4_0 patch embedding + transformer blocks
tokenizer.json ~12 MB Qwen3.5 BPE tokenizer (248,320 vocab)
tokenizer_config.json ~16 KB Special-token IDs, chat-template metadata
chat_template.jinja ~8 KB Standalone chat template
loras/quotes/adapter_config.json ~400 B LoRA manifest (peft_type: LORA, r=16, alpha=16)
loras/quotes/adapter_model.safetensors ~138 MB LoRA weights — fine-tuned on the published quotes dataset

Total cold-load on first visit: ~2.9 GB (LM + vision) or ~2.6 GB (LM only). Subsequent visits are instant — IndexedDB-cached.

Quantization quality

Measured against the BF16 base on a 129-task suite (24 VQA, 20 reasoning, 5 captions, 80 POPE-adversarial public VQA), greedy decoding @128 tokens. The Q4 weights are dequantized shader-exactly into the base layout, so the only difference is weight precision:

Metric BF16 base This Q4
VQA accuracy 0.833 0.917
Reasoning accuracy 0.65 0.65
POPE-adversarial accuracy 0.525 0.525
Caption F1 (lexical) 0.172 0.184

Grade agreement with base: 0.95; mean lexical similarity of Q4 outputs to base outputs: 0.71 — the Q4 model's greedy decode tracks the BF16 base almost token-for-token. (Absolute scores are depressed on both sides by this tier's verbose step-by-step answer style at the fixed 128-token budget; the base-vs-Q4 delta is the meaningful number.)

Format note

These files target mentria-engine's specific Q4_0 safetensors layout. They are produced by an offline conversion pipeline from the upstream BF16 checkpoint (with MSE-optimal per-block Q4 scales) and are intended for use by mentria-engine specifically.

Model details

Base model Qwen/Qwen3.5-4B
Architecture Hybrid Gated-DeltaNet (24 layers, asymmetric: 16 key heads / 32 value heads) + GQA Attention (8 layers, 16Q/4KV) + SwiGLU MLP, 32 layers, 4B params, hidden 2560
Quantization Q4_0 (group size 32, F16 scale per K-block, MSE-optimal scales)
Vision tower 24-layer ViT, 1024 hidden, patch 16, spatial merge 2, projects to 2560-dim text embeddings
Tokenizer Qwen3.5 BPE, 248,320 vocab
License Apache 2.0

LoRA: loras/quotes/

Hot-swappable fine-tune for the motivational-quote use case. On this tier the DeltaNet value/gate projections are per-VALUE-head (32 heads), which mentria-engine's asymmetric LoRA dispatch handles natively.

Training data mentriaai/motivational-quotes — 581 hand-curated original quotes, diversity-engineered
Hyperparameters rank 16, alpha 16 (gentle steer for the strong base), dropout 0.05, LR 1e-4 cosine-decay, 400-iter checkpoint, AdamW, completion-format with prompt masking
Target modules mlp.{down,gate,up}_proj, self_attn.{q,k,v,o,a,b,g}_proj (full-coverage fine-tune)

About mentria.ai

Mentria is a creative studio for tools, experiments, and visual transmissions. All tools run locally in your browser with zero server dependency.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for mentriaai/Qwen3.5-4B-mentria

Finetuned
Qwen/Qwen3.5-4B
Quantized
(312)
this model