Codestral, the brand-new code model from Mistral AI, has arrived to the delight (or not) of developers! Created by the Mistral AI team and launched yesterday, Codestral is an open code generation model specifically designed for code generation tasks. It supports over 80 programming languages, including essentials like Python, Java, C++, JavaScript, and even the venerable Fortran.
If you’re struggling with a function, it will help you complete it in no time, and if you need tests, this expert will generate them for you while you take a coffee break. With its fill-in-the-middle capability, it can complete any piece of code, even the most complex ones.
In terms of performance, it sets new records. It’s a model with 22 billion parameters, but thanks to its 32,000-token context window, it outperforms other models on numerous benchmarks, including RepoBench for long-range code generation. On specific tests like HumanEval, MBPP for Python, and Spider for SQL, Codestral also shows impressive performance, outpacing GPT-4-Turbo and GPT-3.5.
This model is available on HuggingFace, where you can download it, and with the Mistral AI API, a few lines of code are enough to integrate it into your project. Fans of VSCode and JetBrains will be pleased as the Continue.dev and Tabnine plugins have also integrated this model, allowing you to code and interact with it directly in your favorite IDE. You can also access it via the Mistral Chat.
Personally, I use Claude 3 Opus and GPT-4o for my code, but I will start testing this one as well. In any case, if you’re interested, check out the official documentation of Mistral AI here.