Codestral is Mistral AI's first model for code-generation tasks
Mistral AI recently launched Codestral, a 22B-parameter model focused at helping developers with code-generation tasks, and shown to outperform models with more significant system requirements, including CodeLlama 70B, Llama 3 70B, and DeepSeek Coder 33B.
After making the news for the speculation surrounding its $600 million raise at a $6 billion valuation, Mistral has announced the release of Codestral, its new code model. According to the company, Codestral handles English and over 80 coding languages, ranging from the general Python, Java, C, C++, JavaScript, and Bash, to the more specific Swift and Fortran. These capabilities ensure that Codestral can assist developers in various contexts by completing code functions, writing tests, and completing partial code using a fill-in-the-middle mechanism.
Codestral was introduced as an open-weight model, so it is not fully open-source. Launched in parallel with Codestral, the Mistral AI Non-Production License, a license detailing what counts as non-commercial and research supporting uses of Mistral AI's products. Codestral is the first model to be launched under this new license, although Mistral AI has stated it wil continue to release models and code under the Apache 2.0 license as it continues to consolidate a two families of products, one for each license.
Codestral is a 22B-parameter model, which means it places some hefty demands on the systems that can run it, thus making it impractical for some contexts. This may be offset by the fact it ranks highly in benchmarking tests against models with even higher system requirements, including CodeLlama 70B, Llama 3 70B, and DeepSeek Coder 33B. The fact that Codestral can perform comparably or even better that models more than twice its size represents a commendable breakthrough in the code generation space. None of this entails, however, that Codestral is different to any other AI coding tool. As always, it should be used with care, especially in view of recent studies putting the alleged productivity boost of coding assistants such as GitHub Copilot into perspective.
Codestral is available for download at Hugging Face; via the dedicated codestral.mistral.ai endpoint (free for the next 8 weeks); via the usual Mistral AI endpoint; through Le Chat, Mistral AI's conversational interface; and through the LlamaIndex and LangChain integrations, as well as the VSCode/JetBrains integration.
