Home
Moore Threads Open-Sources MusaCoder, China's First Full-Stack Self-Developed Code AI Model

China's AI computing power sector has achieved a major milestone. Recently, Molyneux officially released and open-sourced MusaCoder, a code large model designed specifically for generating GPU-level operators. It is the first open-source code model in the industry to have its full training and verification process completed on a fully functional domestic GPU platform.
From a technical perspective, the development of MusaCoder represents a significant step forward for China's computing power ecosystem. The entire post-training process was carried out on the "Kuao" computing cluster powered by MTT S5000 GPUs. This demonstrates that domestic hardware can reliably and efficiently support the full-chain development of complex large models, offering the industry a complete blueprint spanning from underlying hardware to upper-level model training.
Performance-wise, MusaCoder proves highly competitive. In the widely recognized KernelBench strict evaluation, the MusaCoder-27B-RL model achieved outstanding results: an Overall Pass rate of 93.2% and an average score of 88.60%. These results show it has surpassed several internationally renowned SOTA code models, including Claude Opus 4.7, DeepSeek-V4 Pro, GLM-5.1, and Kimi K2.6, placing it among the top tier in the industry.
This open-source release is not just a technical milestone for Molyneux in the model space, but also a key move to strengthen the domestic computing power ecosystem. In recent years, Molyneux has been steadily deepening its work on the underlying ecosystem, completing adaptations for major models like DeepSeek, Qwen, and MiniMax, and releasing open-source operator development tools along with other supporting solutions. With the official launch of MusaCoder, developers can more easily leverage domestic computing power to speed up operator development and model training, further unlocking the potential of full-featured domestic GPUs.
Industry analysts note that code models act as the core engine for AI development, making their performance and autonomy critical. Molyneux's MusaCoder, built through a full-stack training approach, offers a more independent tool option for domestic AI research and development, which is essential for laying a stronger foundation for homegrown AI technology.
Related article
MiniMax 2.5 Tops Global Usage With Over $150M Monthly Revenue; M3 Flagship Version Set for H1 Release
Chinese large language models are generating intense momentum across the global developer ecosystem, driven by their reputation for cost-efficiency and engineering excellence.As of March 3, 2026, data from OpenRouter, the world’s leading API aggregat
Meta's Muse outpaces ChatGPT in early mobile launch
New data suggests Meta’s latest AI app, Muse, could become the social giant’s next major success. According to recent estimates from Appfigures, Muse has been downloaded on mobile devices more times in its first twelve days than ChatGPT did in the tw
TechCrunch Disrupt 2026: Clay’s Kareem Amin on the rise of the GTM engineer
Two years ago, GTM engineers were virtually nonexistent in most organizations. Today, the role is emerging across the tech sector, with professionals leveraging AI and automation to construct revenue engines that previously demanded multiple hires, d
Related Special Topic Recommendations
Comments (0)
0/500

China's AI computing power sector has achieved a major milestone. Recently,
From a technical perspective, the development of MusaCoder represents a significant step forward for China's computing power ecosystem. The entire post-training process was carried out on the "Kuao" computing cluster powered by
Performance-wise, MusaCoder proves highly competitive. In the widely recognized KernelBench strict evaluation, the MusaCoder-27B-RL model achieved outstanding results: an Overall Pass rate of 93.2% and an average score of 88.60%. These results show it has surpassed several internationally renowned SOTA code models, including Claude Opus 4.7, DeepSeek-V4 Pro, GLM-5.1, and Kimi K2.6, placing it among the top tier in the industry.
This open-source release is not just a technical milestone for Molyneux in the model space, but also a key move to strengthen the domestic computing power ecosystem. In recent years, Molyneux has been steadily deepening its work on the underlying ecosystem, completing adaptations for major models like DeepSeek, Qwen, and MiniMax, and releasing open-source operator development tools along with other supporting solutions. With the official launch of MusaCoder, developers can more easily leverage domestic computing power to speed up operator development and model training, further unlocking the potential of full-featured domestic GPUs.
Industry analysts note that code models act as the core engine for AI development, making their performance and autonomy critical. Molyneux's MusaCoder, built through a full-stack training approach, offers a more independent tool option for domestic AI research and development, which is essential for laying a stronger foundation for homegrown AI technology.
MiniMax 2.5 Tops Global Usage With Over $150M Monthly Revenue; M3 Flagship Version Set for H1 Release
Chinese large language models are generating intense momentum across the global developer ecosystem, driven by their reputation for cost-efficiency and engineering excellence.As of March 3, 2026, data from OpenRouter, the world’s leading API aggregat
Meta's Muse outpaces ChatGPT in early mobile launch
New data suggests Meta’s latest AI app, Muse, could become the social giant’s next major success. According to recent estimates from Appfigures, Muse has been downloaded on mobile devices more times in its first twelve days than ChatGPT did in the tw
TechCrunch Disrupt 2026: Clay’s Kareem Amin on the rise of the GTM engineer
Two years ago, GTM engineers were virtually nonexistent in most organizations. Today, the role is emerging across the tech sector, with professionals leveraging AI and automation to construct revenue engines that previously demanded multiple hires, d











