CodeShell is a series of open-source code large language models developed by Peking University’s Knowledge Computing Lab together with the AI team at Sichuan Tianfu Bank. With about 1,600 stars on GitHub, its flagship 7-billion-parameter base model was trained on 500 billion tokens and offers an 8,192-token context window.
CodeShell was built as a strong foundation model for code: at release it reported the best results among models of its size on the HumanEval and MBPP code generation benchmarks. The family focuses on multilingual code generation, with checkpoints released through Hugging Face and ModelScope for fine-tuning and downstream research. The GitHub repository collects the model cards, usage guidance and links to the weights, while the models themselves are distributed through the model hubs. As an academic-industry collaboration, CodeShell is aimed at researchers and engineers who want a capable mid-sized code model they can adapt and study, rather than a hosted coding assistant product.
The release family also includes additional checkpoints listed from the same repository and model hub pages, so teams can pick the variant that best matches their hardware budget and accuracy requirements.
CodeShell’s models and code are free and openly released through GitHub, Hugging Face and ModelScope. There is no hosted API or paid tier from the project; you download the weights and run them on your own hardware.
Researchers and engineers who want a mid-sized open code model to fine-tune or evaluate, especially teams working on multilingual code generation outside the largest commercial model ecosystems. It is a base model release, so expect to build your own tooling around it.