Mistral AI releases Devstral 2 and the Mistral Vibe CLI

Background

Coding has become one of the main uses of large language models, and model makers increasingly pair a model with a tool that plugs into a developer's terminal or editor. Mistral AI's release on 9 December 2025 followed that pattern: two coding models plus a command-line agent to drive them.^[1]

What was released

Devstral 2 is a dense model with 123B parameters. Alongside it, Mistral released Devstral Small 2, a 24B model in the same family. Both are coding models, aimed at software engineering tasks rather than general chat. Both have 256K context windows, so they can take in a large amount of code and instruction text at once — useful when a model has to reason over several files of a project rather than a single snippet. The two differ in licensing: Devstral 2 ships under a modified MIT license, while Devstral Small 2 uses Apache 2.0.^[1]

Benchmark results

Mistral reported scores on SWE-bench Verified, a benchmark built from real software issues that a model must resolve. Such benchmarks are used because they test whether a model can change an existing codebase rather than answer a question in isolation. Devstral 2 scored 72.2%, and Devstral Small 2 scored 68.0%. The gap is narrow given the difference in size: the 24B model trails the 123B model by a little over four points.^[1]

Mistral Vibe

Mistral also released Mistral Vibe, an open-source command-line coding agent under the Apache 2.0 license. Mistral describes it as an open-source command-line coding assistant powered by Devstral. The agent is the interface — it runs in the terminal, takes instructions and works on code — while the Devstral models supply the underlying capability.^[1]

Availability

At launch, Devstral 2 was free through the Mistral API, with no end date given, so developers could try the larger model without a stated cutoff.^[1]

Why it matters

The release puts a larger open model and a small one side by side, with the small model close behind on a demanding coding benchmark — a combination that matters to teams choosing between running a model locally and calling one over an API. Shipping a command-line agent under a permissive license also gives developers a ready-made way to use the models, instead of leaving them to build their own tooling.^[1]