Codestral was introduced by Mistral AI as their first model purpose-built for code generation, trained on a diverse corpus spanning more than eighty programming languages including widely used ones like Python, Java, C, C++, JavaScript, and Bash as well as more specialized languages such as Swift and Fortran. The model is exposed through a shared instruction and completion API endpoint, allowing developers to both prompt it with natural-language instructions and use a fill-in-the-middle mechanism to complete partial functions, write tests, or extend existing snippets. This dual endpoint design makes Codestral flexible for interactive development workflows where both conversational guidance and direct code in-filling are useful.

Positioned as a 22B-parameter model, Codestral was designed to offer a favorable balance between performance and latency relative to earlier code-specific alternatives, and it was reported to set a new standard on long-range code evaluations such as RepoBench thanks to a 32k-token context window that exceeded the 4k to 16k ranges of competing models at the time. Google Cloud subsequently became the first hyperscaler to host Codestral as a fully-managed service through Vertex AI Model Garden, broadening access for teams that prefer managed inference over self-hosting. Codestral's combination of broad language coverage, fill-in-the-middle completion, and a generous context window makes it well suited for tasks such as code completion, test generation, documentation drafting, and building developer-assistant applications across multiple language ecosystems.