Moore Threads announces open-sourcing of MusaCoder: the first domestically developed, full-featured GPU full-stack training code for large-scale models.
Moore Threads announced the official release and open source of MusaCoder, a large-scale dedicated code model for generating low-level GPU operators.
Moore Threads stated that MusaCoder is the industry's first open-source large model that has completed the entire training and verification process based on a domestically produced GPU computing platform. Its complete post-training process was completed on the Kua'e Intelligent Computing Cluster built on MTT S5000.
According to reports, MusaCoder includes two parameter scales, 9B and 27B, and is mainly designed for GPU low-level operator generation tasks, with a focus on supporting the automatic generation of high-performance CUDA/MUSA native kernel code from PyTorch standard operators.
This capability can lower the barrier for developers to write low-level GPU operators by hand and improve the efficiency of code generation, verification and optimization in high-performance GPU computing scenarios.
In terms of performance, in the KernelBench test, the MusaCoder-27B-RL achieved Overall Pass@8 93.2% and Avg.@8 88.60%, surpassing mainstream state-of-the-art code models such as Claude Opus 4.7, DeepSeek-V4 Pro, GLM-5.1, and Kimi K2.6, reaching the current industry-leading level.
According to Moore Threads, MusaCoder's full-stack training and verification process, including SFT (supervised fine-tuning), RFT (rejection sampling fine-tuning), RL (reinforcement learning), asynchronous rollout, online compilation and execution verification, and reward calculation, is all completed using the Kua'e Intelligent Computing Cluster built on MTT S5000.
