Moore Threads has released and open‑sourced MusaCoder, a code‑focused large language model optimized for GPU‑level operator generation. It is the first open‑source code LLM trained and validated end‑to‑end on a domestic full‑stack GPU compute base; all post‑training was completed on the Kua'e intelligent compute cluster built on MTTS 5000. In KernelBench's strict evaluation, MusaCoder‑27B‑RL recorded OverallPass@893.2% and Avg.@888.60%, outperforming Claude OPUS 4.7, DeepSeek‑V4 Pro, GLM‑5.1 and

Moore Threads has released and open‑sourced MusaCoder, a code‑focused large language model optimized for GPU‑level operator generation. It is the first open‑source code LLM trained and validated end‑to‑end on a domestic full‑stack GPU compute base; all post‑training was completed on the Kua'e intelligent compute cluster built on MTTS 5000. In KernelBench's strict evaluation, MusaCoder‑27B‑RL recorded OverallPass@893.2% and Avg.@888.60%, outperforming Claude OPUS 4.7, DeepSeek‑V4 Pro, GLM‑5.1 and Kimi K2.6 and reaching industry‑leading performance.