Codestral 2

Codestral 2 is a Mistral-built code-generation model engineered around high-precision fill-in-the-middle completion, exposed through a shared instruction-and-completion endpoint so editors and agents can drive both interactive suggestions and longer generative tasks from a single interface. It is positioned inside Google's Vertex AI as a third-party partner model, where Mistral remains the original creator while the platform handles hosting, region availability, and quota management. Practically, the model is aimed at software developers and tooling builders who need reliable in-editor code writing, translation between languages, and conversational explanations of code without leaving their development environment.

Relative to the prior Codestral 25.01 generation, Vertex-documented figures describe a 30 percent lift in accepted completions, about 10 percent more code retained after each suggestion, and roughly 50 percent fewer runaway generations, which together translate into more trustworthy longer edits and fewer wasted keystrokes during interactive coding. The model is tuned for both short and long-context fill-in-the-middle scenarios, and supports code generation, summarization, documentation, refactoring, bug fixing, and test-case authoring on top of its core FIM strength. These characteristics make Codestral 2 a focused fit for IDE plugins, code review assistants, and developer-facing agents that depend on precise, contained edits rather than open-ended chat.