Jacqkues/d3-lite-decisionjs

d3-lite for decision.js (q8, WebGPU)

vLLM Semantic Router's d3-lite (Decision 3.0, built on Qwen3.5-0.8B), converted for decision.js: typed decisions (yes/no, choice, score) over any text or JSON, in the browser on WebGPU, in one forward pass per question and with no token generated.

Try it: jacqkues.github.io/decision/playground

decision.js loads it from this repo's resolve/main/ URL (Decision.load("https://huggingface.co/Jacqkues/d3-lite-decisionjs/resolve/main/")).

Files

• manifest.json: the model config and where each tensor lives in weights.bin (decision.js format d1.js/1);
• weights.bin (803 MB): the language model, read with HTTP Range requests;
• tokenizer.json, tokenizer_config.json: the original tokenizer;
• reference.json: the official runtime's probabilities on decision.js's parity requests;
• LICENSE, NOTICE: the original model's license and notice.

Changes from the original

This is a modified copy of vllm-sr/d3-lite (main, as of 2026-10-12), made with decision.js's scripts/export-d3.py:

• the linear weights quantized to 8 bits in blocks of 32 (absmax scale in f16) and the embedding likewise;
• the attention's q / k / v / gate and the MLP's gate / up projections fused; the DeltaNet in-projections fused;
• the zero-centred RMSNorm weights folded (1 + w); norms, conv taps and gates in f16; the code readout in f32;
• the vision tower left out (text only).

On decision.js's parity set it gives the same answer as the original BF16 model on 56 / 56 questions; on a 465-question sample of the Decision Index it scores 77.8% against 77.6% for the official runtime.

License

Apache-2.0, as the original (see LICENSE and NOTICE). d3-lite is © 2026 vLLM Semantic Router Team, built on Qwen/Qwen3.5-0.8B (Apache-2.0).