Qwen3.5-0.8B-mentria
Qwen3.5-0.8B 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 default tier of the mentria model ladder (0.8B → 2B → 4B) — it runs on every WebGPU-capable device, including phones; capable devices can be served a larger tier.
Try it live:
- mentria.ai/tools/ai-chat — text + vision chat
- mentria.ai/tools/quote — quote generator (uses the bundled LoRA)
Files
| Path | Size | Purpose |
|---|---|---|
qwen3.5-0.8b-q4-tied.safetensors |
~450 MB | Language model (default) — Q4_0 weights (MSE-optimal per-block scale) with tied embedding/lm_head (the embedding table is shared with the output projection, so it ships once). This is the bundle mentria-engine loads (allowTiedEmbed). |
qwen3.5-0.8b-q4.safetensors |
~602 MB | Language model (legacy, untied) — separate embed_tokens + lm_head. Kept for compatibility. |
qwen3.5-0.8b-vl-q4.safetensors |
~60 MB | Vision tower — Q4_0 patch embedding + transformer blocks (MSE-optimal scales) |
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=32) |
loras/quotes/adapter_model.safetensors |
~44 MB | LoRA weights — fine-tuned on curated motivational quotes |
Total cold-load on first visit: ~510 MB (tied LM + vision) or ~450 MB (tied 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 (MSE scales) |
|---|---|---|
| VQA accuracy | 0.792 | 0.792 |
| Reasoning accuracy | 0.60 | 0.65 |
| POPE-adversarial accuracy | 0.850 | 0.825 |
| Caption F1 (lexical) | 0.666 | 0.693 |
Grade agreement with base: 0.895. The MSE-optimal per-block scales recover ~5.6% of quantization weight error versus plain min-max Q4_0 and moved every measured metric toward (or past) the base.
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/Qwen3.5-0.8B |
| Architecture | Hybrid DeltaNet (18 layers) + GQA Attention (6 layers) + SwiGLU MLP, 24 layers, 0.8B params |
| Quantization | Q4_0 (group size 32, F16 scale per K-block, MSE-optimal scales) |
| Vision tower | 12-layer ViT, 768 hidden, patch 16, spatial merge 2, projects to 1024-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.
| Training data | Curated motivational quotes in chat format. A newer diversity-engineered dataset is published at mentriaai/motivational-quotes and ships on the 2B/4B tiers; at 0.8B scale its diversity degrades coherence — small bases need a rigid template style, so this tier deliberately keeps its original adapter |
| Hyperparameters | rank 16, alpha 32, dropout 0.05, LR 2e-4 cosine-decay, AdamW, prompt-masked |
| 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.