China's artificial intelligence computing power sector has made a significant breakthrough. Recently, Molyneux officially released and open-sourced a code large model specifically designed for generating GPU-level operators - MusaCoder. This is also the first open-source code model in the industry that has completed the entire training and verification process based on a fully functional domestic GPU platform.

In terms of technical implementation, the creation of MusaCoder marks an important advancement in China's computing power ecosystem. The entire post-training (Post-training) process of this model was completed on the "Kuao" computing cluster built using MTT S5000 GPUs. This achievement proves the reliability and efficiency of domestic hardware in supporting complex large model full-chain development tasks, providing the industry with a complete reference pattern from the underlying hardware to the upper-level model training.

In terms of performance, MusaCoder demonstrates strong competitiveness. In the industry-recognized KernelBench strict evaluation, the MusaCoder-27B-RL model achieved remarkable results: its Overall Pass rate reached 93.2%, with an average score of 88.60%. This test result shows that the model has surpassed multiple internationally renowned SOTA (State-of-the-art) code models, including Claude Opus 4.7, DeepSeek-V4 Pro, GLM-5.1, and Kimi K2.6, securing its position in the top tier of the industry.

This open source initiative is not only a technical accumulation of Molyneux in the model field, but also a key measure in its efforts to improve the domestic computing power ecosystem. In recent years, Molyneux has been continuously deepening its work on the underlying ecosystem, completing the adaptation of multiple mainstream large models such as DeepSeek, Qwen, and MiniMax, and launching open-source operator development tools and other supporting solutions. With the official release of MusaCoder, developers can more conveniently utilize the domestic computing power base to accelerate operator development and model training processes, further unleashing the computing potential of fully functional domestic GPUs.

Industry analysts point out that code models serve as the core "engine" for AI development, and their performance and autonomy are crucial. Molyneux's MusaCoder, developed through a full-stack training path, provides a more autonomous tool choice for domestic AI research and development, which is of great significance for building a more solid foundation for domestic artificial intelligence technology.