nakas/chessgpt-coach-4b-v1-q4f16_1-MLC

chessgpt coach 4B (q4f16_1, MLC / WebLLM)

Qwen3-4B fine-tuned (LoRA, merged) to explain chess mistakes for chessgpt: it narrates engine-verified facts (Stockfish lines and program-detected motifs) at the player's rating level. Runs in the browser with WebLLM using the prebuilt Qwen3-4B q4f16_1 WebGPU library.

Weights are WebLLM's own mlc-ai/Qwen3-4B-q4f16_1-MLC build with the attention and MLP projections replaced by the fine-tuned ones, quantized the same way (checked byte-for-byte against the official build on the base model). Base model license: Apache-2.0.