Mistral AI and All Hands AI release Devstral

Background

An agentic model is not used to answer a single question but as the engine of a program that works in steps: it reads files, proposes changes, runs commands, looks at the results and decides what to do next. Software engineering became one of the earliest testing grounds for this pattern, because the work is naturally iterative and because the outcome can be checked — a proposed patch either helps resolve an issue or it does not.

What was released

On 21 May 2025, Mistral AI released Devstral, an open agentic coding model built with All Hands AI ^[1]. Mistral described it as an agentic LLM for software engineering tasks, trained to solve real GitHub issues ^[1]. That description places the model in the part of the field concerned less with generating snippets in a chat window and more with carrying out multi-step work inside a codebase, where the model has to locate the relevant code, understand what is wrong, and produce a change.

The result it was measured on

On SWE-Bench Verified, a benchmark used to test whether a model can resolve real software issues, Devstral scored 46.8% ^[1]. Mistral said that the model outperforms all open-source models on that benchmark by a large margin ^[1]. Benchmark comparisons of this kind are the releasing company's own reading of the numbers, and they are the usual way coding models are introduced: a score on a shared task, set beside the scores of other models.

Licence, size and price

Devstral is released under the Apache 2.0 licence ^[1], which allows the model to be used and modified without the restrictions of a closed release. It also runs on a single RTX 4090 ^[1]. Mistral said the model is light enough to run on a Mac with 32GB RAM ^[1], a claim that matters for people who want to try an agentic coding model on hardware they already own rather than through a remote service.

For those who do prefer a service, the model is available on Mistral's API as devstral-small-2505, at $0.1 per million input tokens and $0.3 per million output tokens ^[1].

Why it matters

Three properties of the release sit together. The model is openly licensed, so its weights can be run and adapted outside the company that made it ^[1]. It is small enough to fit on a single consumer graphics card and on a laptop with 32GB of memory ^[1]. And it is built for agentic use on software tasks rather than for general conversation ^[1]. Each of these has been available separately in earlier coding models; the combination is what defines Devstral's place in the field.

A coding agent that can be run locally also changes who can experiment with one. When the model fits on hardware a developer already owns, testing an agent on a private codebase does not require sending that code to an external API. When it is offered on an API at a published per-token price, the same model can be used at larger scale without local hardware ^[1].

The benchmark claim is the part most likely to be tested by others. Mistral's statement that Devstral outperforms all open-source models on SWE-Bench Verified by a large margin is a claim about a moving target: open models are compared against each other continuously, and a lead measured on one benchmark on one day is a snapshot rather than a settled ranking ^[1]. What the release establishes more durably is the shape of the offering — an openly licensed, small-footprint, agent-oriented coding model, made with a partner specialised in agent tooling, and priced for use through an API ^[1].