Devstral

Devstral is Mistral AI's answer to the question: what does a model look like when it is purpose-built for coding agents rather than adapted from a general-purpose model? Instead of training a large model on everything and then fine-tuning it for code, Devstral is optimized from the ground up for agentic programming tasks: reading and navigating codebases, calling tools in sequence, planning multi-step implementations, and recovering from errors mid-task. Its 256,000-token context window (roughly equivalent to a few hundred pages of dense code) allows it to hold large codebases in memory during a single session.

Devstral 2 is the current version in mid-2026, appearing on model leaderboards inside tools like Kilo Code and described by Mistral as a state-of-the-art open-source model for agentic coding. The open-weights release (meaning the model parameters can be downloaded and run locally rather than only accessed via an API) is significant for teams with data-sovereignty requirements or cost constraints, since self-hosting eliminates per-token API costs after the initial infrastructure setup.

For most builders, Devstral sits in Specialist territory: it is the kind of model choice that comes up when you are benchmarking your agentic coding pipeline against specific metrics, comparing open-weight alternatives to Claude or GPT for a self-hosted deployment, or building a product where you need to customize the model's behavior through fine-tuning (additional training on your own data). If you are just getting started with AI coding agents, Claude Code, Cursor, or Kilo Code with a cloud model are more practical starting points.