TMTPOST — Graphics processor designer Moore Threads Technology Co., Ltd. has officially released and open-sourced a specialized large language model called MusaCoder, designed to automatically generate low-level compute kernels for graphics processing hardware, according to a technical announcement on Wednesday. Marking an industry first, the development team completed the entire training and verification pipeline for this open-source coding model strictly on a domestic hardware foundation. Engineers executed the full post-training phase utilizing the company's proprietary KUAE Intelligent Computing Cluster, which is built entirely on MTT S5000 GPU nodes. The release marks a significant milestone in building an independent software ecosystem for domestic semiconductor hardware. By deploying custom AI models to automate the complex programming of low-level mathematical operations, the chip designer aims to accelerate software optimization and lower development barriers for engineers writing applications for its silicon architecture.