manishkumar2101114/qwen3.6-35b-a3b-tool-hindi-merged-fp8

Qwen3.6-35B-A3B tool-Hindi-merged FP8

Base: Qwen/Qwen3.6-35B-A3B-FP8 (same FP8 block format/scales convention). Adapter: qwen3.6-35b-a3b-tool-hindi-adapter (LoRA r=32, 3 epochs, 1693 dialogues incl. 178 Hindi-number drills). Train loss 0.2225 / eval 0.1683. Merged in BF16, block-requantized to FP8 e4m3/128. Chat template: qwen3_template333.jinja. Serve with sglang / vllm / transformers.