What is Devstral?
Devstral is an open-weight coding model from Mistral AI, the European lab, built specifically for agentic software engineering. Where a general model answers questions about code, Devstral is tuned to act: to run inside an agent harness, use tools, edit across multiple files, and resolve real repository issues.
It was developed with the open-source agent project OpenHands and is designed to plug into that kind of scaffolding. The lineup includes a smaller model you can run on modest hardware and a more capable variant, with permissive open weights so you can self-host.
Devstral is a strong fit for teams that want a capable coding agent under their own control: open weights for data residency, European governance, and the freedom to run it on your own infrastructure or through Mistral's API.
What it's best for
- Powering coding agents that edit across files and resolve real issues.
- Self-hosting an agentic coding model for data residency and control.
- Teams wanting a European, openly licensed alternative to closed coding models.
- Running inside open agent scaffolds such as OpenHands.
Where it falls short
- General chat or non-coding work, where a broader model fits better.
- Teams wanting a finished, hosted agent product rather than a model to build with.
- Running the larger variant without enough GPU memory to host it.
Ways in
Use Devstral through Mistral's API, or download the open weights and self-host. The smaller model is light enough for a single capable GPU; the larger one needs more.
Pair it with an agent harness such as OpenHands rather than treating it as a chatbot; Devstral is built to drive tools and edits, not just to talk about code.
Getting the most out of it
Give the agent the repository context and a clear task with a verification step, and let the model plan and execute the edits.
For data-sensitive work, self-host the weights so the code never leaves your environment, and confirm the license covers your use.
What Devstral costs
Approximate, in USD, as of June 2026. Prices change often. Confirm on the official site before you rely on them.
Open weights
$0 (self-host)
Devstral weights are released under a permissive open license; you pay only your own compute. Confirm the exact license on the official site.
API
Low, usage-based
Priced per token through Mistral's platform. Confirm current rates on the official site.
Example prompts
Copy these into Devstral as starting points, then adapt them to your task.
Here is the repository and an open issue. Reproduce the problem, implement a fix across the files involved, and make the tests pass.
Rename this concept throughout the codebase, update every reference and import, and confirm the build still succeeds.
List the hardware needed to run the smaller Devstral model locally, and the steps to serve it behind an OpenAI-compatible endpoint.
Devstral
common questions.
Direct answers to the questions we get asked the most. If yours isn't covered, write to the team.