Codestral 25.08 for business

Codestral 25.08 is Mistral's current Codestral model, released in July 2025. It is built for low-latency code completion and generation, including fill-in-the-middle, where the model completes code between existing lines rather than only at the end. It supports tool calling and structured outputs, and the official card lists a 128K token context window.

Speed matters for code completion. A suggestion that arrives while the developer is still typing is useful. One that arrives two seconds later is ignored. That is why smaller, faster coding models still have a place next to large models such as Claude Opus or GPT-6 Sol.

For larger tasks, such as changing a feature across many files, we would use a stronger model in an agent setup and keep Codestral for completion. In our own work, every AI-generated change is reviewed by an engineer. See AI-accelerated development.

Codestral is listed under Mistral's commercial license rather than as open weights, so it is used through the API. If you need a coding model you can host yourself, look at Qwen3-Coder or Mistral Small 4.

When building completion into an internal tool, we would measure acceptance rate: how often developers keep the suggestion. It is the simplest and most honest measure of whether a coding model helps, and it is easy to track over time. See AI copilots.