MooreThreads releases MusaCoder-27B, a 27B-parameter code generation model
China's GPU maker MooreThreads enters AI code generation with a 27B open-source model...
MooreThreads, the Chinese company known for developing the MUSA GPU architecture (an alternative to CUDA), has released a large language model focused on code generation. The model, named MusaCoder-27B, boasts 27 billion parameters and is available on Hugging Face under an open-source license. According to the associated arXiv paper (2606.04847), the model was trained on a diverse corpus of programming languages, including Python, C++, Java, and Chinese-specific code repositories. It reportedly achieves competitive performance against similar-sized models like Code Llama 34B and StarCoder 15B on standard benchmarks such as HumanEval and MBPP.
The release marks a significant step for MooreThreads in the AI ecosystem, demonstrating that their hardware can effectively train large-scale generative models. The MusaCoder-27B is designed to run inference on MUSA GPUs, but also supports conversion to other frameworks via ONNX. Developers can use the model for code completion, bug fixing, and documentation generation. Being open-source, it allows researchers to fine-tune it for domain-specific tasks, such as embedded systems or RISC-V programming, which are strategic areas for MooreThreads.
- 27-billion parameter code generation model released by MooreThreads on Hugging Face
- Trained on a mixture of general and Chinese-specific code repositories
- Compatible with MooreThreads' MUSA GPUs and supports ONNX conversion for cross-platform use
Why It Matters
MusaCoder-27B opens Chinese hardware-optimized AI code generation to the open-source community, challenging Western models.