Moore Threads open-sources MusaCoder, a code LLM fully trained on domestic GPUs

摩尔线程开源 MusaCoder 代码大模型，9B/27B 参数基于国产 GPU 全链路训练

Moore Threads released MusaCoder, a code model for GPU kernel generation, in 9B and 27B sizes. The full post-training pipeline ran on a domestic MTT S5000 cluster. It auto-generates high-performance CUDA/MUSA kernels from PyTorch ops. On KernelBench, the 27B RL version hits 93.2% Overall Pass@8, beating Claude Opus 4.7 and DeepSeek-V4 Pro per the official report. Models and paper are public.

Why it matters: Moore Threads open-sourced a code model trained end-to-end on its own MTT S5000 GPUs, with 9B and 27B variants targeting CUDA/MUSA kernel generation. The story's edge is the full domestic-hardware loop — chip, training, and model output all in-house, not a wrapper. Score stays...