Moore Threads open-sources GPU code model MusaCoder, outperforming mainstream models

Moore Threads has released and open-sourced MusaCoder, a specialized code model for GPU kernel generation, available in 9B and 27B parameter sizes. It's the first open-source code model fully trained and verified on domestic GPU hardware, with the entire post-training process completed on the Kuae AI computing cluster built with MTT S5000. The model focuses on automatically generating high-performance CUDA/MUSA kernel code from PyTorch standard operators, lowering the barrier for developers writing GPU kernels. In KernelBench tests, MusaCoder-27B-RL achieved Overall Pass@8 of 93.2% and Avg.@8 of 88.60%, outperforming models like Claude Opus 4.7 and DeepSeek-V4 Pro, reaching industry-leading performance.