Codestral was Mistral AI's first dedicated code model, launched in May 2024 at 22 billion parameters, trained across more than 80 programming languages including Python, Java, C, C++, JavaScript, Bash, Swift, and Fortran. It supports both instruction-based completion and fill-in-the-middle workflows, where a model fills in missing code between existing lines rather than only appending to the end, and it performed well on benchmarks like RepoBench, HumanEval FIM, MBPP, and Spider for SQL generation.
A 2025 update, Codestral 25.01, roughly doubled inference speed through an improved architecture and tokenizer, strengthening its position specifically on fill-in-the-middle tasks where speed matters more than in a typical chat interaction. Its context window has grown to 256,000 tokens, letting it reason over much larger spans of a codebase at once than its original release could.
Mistral relicensed Codestral to Apache 2.0 in April 2026, after initially shipping it under a restrictive non-production license, which opened it up to genuine commercial use and self-hosting without licensing fees, a meaningful shift from its early positioning as a research and evaluation model.