In-browser (WebGPU) q4f16_1 build of a Qwen3-1.7B student distilled for structured code review. Runs client-side via WebLLM / MLC-LLM — no server, no data leaves the machine.

What it is

Student: Qwen3-1.7B, LoRA-distilled (forward-KL / TAID, native temperature, response-position).
Teacher: Qwen3-8B (8-bit), targets generated with thinking disabled (clean JSON).
Task: read a numbered code chunk, emit a JSON findings array (category, severity, line, evidence…).
Quality: held-out planted-bug F1 ≈ 0.506 (bf16 eval) — ties the earlier best 3B student (Gemma-3B 0.509) at roughly half the parameters, and beats the old Qwen2.5-Coder-1.5B (0.482).
Format: MLC q4f16_1, ~0.93 GB, context 4096, prefill chunk 1024.

Running in the browser

Pair these weights with WebLLM's prebuilt Qwen3-1.7B model library:

{
  model: "https://huggingface.co/bernardw/qwen3-1.7b-codereview-taid-q4-mlc",
  model_id: "Qwen3-1.7B-CodeReview-TAID-q4f16_1",
  model_lib: "https://raw.githubusercontent.com/mlc-ai/binary-mlc-llm-libs/main/web-llm-models/v0_2_84/base/Qwen3-1.7B-q4f16_1_cs1k-webgpu.wasm",
  overrides: { context_window_size: 4096 }
}

License

Apache-2.0 (Qwen3 base is Apache-2.0; distilled from Qwen3-8B outputs). No Gemma terms apply.