Codestral 22B has a context length of 32K and can interact with code in various environments and projects. It can generate code from scratch, complete coding functions, write tests, and complete partial code. The model covers popular languages like SQL, Python, Java, C, and C++, as well as specific ones like Swift and Fortran. Despite being newly launched, Mistral claims Codestral outperforms existing models in most programming languages. The model is available on Hugging Face under a non-production license and through two API endpoints.
Key takeaways:
- Paris-based AI startup Mistral has launched Codestral, a 22B parameter, open-weight generative AI model that specializes in coding tasks, supporting more than 80 programming languages.
- Codestral is available under a non-commercial license and is claimed to outperform previous models designed for coding tasks, including CodeLlama 70B and Deepseek Coder 33B.
- The model has been trained on a dataset of more than 80 programming languages and can perform a diverse range of coding tasks, including generating code from scratch, completing coding functions, writing tests and completing any partial code.
- Mistral is offering Codestral 22B on Hugging Face under its own non-production license and is also making the model available via two API endpoints: codestral.mistral.ai and api.mistral.ai.