Mistral introduces its Devstral 2 AI models and Vibe CLI

Mistral AI has introduced Devstral 2, a family of artificial intelligence models for coding, along with Mistral Vibe CLI, an assistant that can explore, modify and execute changes in a project from the terminal.

The family comes in two versions: Devstral 2, with 123 billion parameters, and Devstral Small 2, with 24 billion. A parameter is an internal unit the model uses to learn patterns, although having more does not guarantee better results on its own.

An open model for working with code

According to tests published by Mistral, Devstral 2 scores 72.2% on SWE-bench Verified, an evaluation that measures whether an AI can solve real problems drawn from software projects. Devstral Small 2 reaches 68%.

Both models can analyze an entire repository, follow dependencies between files and coordinate changes affecting multiple parts of an application. They can also detect when a solution has failed, fix it and try again.

That enables tasks such as fixing bugs, updating legacy code or preparing changes that would normally require manually reviewing many files. Their 256,000-token context window lets them keep a large amount of project information available during a single session.

Mistral says Devstral 2 can be up to seven times more cost-efficient than Claude Sonnet on real-world tasks. The company also highlights that its models are much smaller than some competitors: Devstral 2 is five times smaller than DeepSeek V3.2 and eight times smaller than Kimi K2, according to its figures.

Size matters because it reduces the resources needed to run the models. Devstral 2 requires at least four H100-class GPUs for a self-hosted installation, while Devstral Small 2 is designed to run on a single consumer GPU and in CPU-only configurations.

Not everything is a win against closed models

In an independent human evaluation using tasks prepared in Cline, Devstral 2 beat DeepSeek R1 in 42.8% of comparisons and lost in 28.6%. However, Claude Sonnet 4.5 remained the preferred model by a significant margin.

The takeaway is less spectacular and more useful: open models have narrowed the gap in coding, but they still do not consistently match the best closed alternatives in every scenario.

Devstral 2 is distributed under a modified MIT license. Devstral Small 2 uses Apache 2.0. Both can be deployed on your own servers and tuned for specific languages or large codebases, although you should review the exact terms of the modified license before incorporating it into a commercial product.

Vibe CLI brings AI to the terminal

Mistral Vibe CLI is an open-source coding agent under the Apache 2.0 license, built to work with Devstral. Instead of merely suggesting a line of code, it can search for files, edit them, run commands and use version control through a conversational interface.

Its features include:

• Automatically analyzing the project structure and its Git status.
• Referencing files with @, running commands with ! and changing settings through slash commands.
• Coordinating changes across multiple files, with a broader view than the currently open editor.
• Saving history, offering autocomplete and allowing custom themes.

It can also run through scripts, configure automatic approval for actions and connect to local models or providers from a config.toml file. It is available as an extension for Zed and can integrate with other environments through the Agent Communication Protocol.

What changes for you

Devstral 2 is available for free through Mistral's API during the initial period. After that, the announced price will be $0.40 per million input tokens and $2 per million output tokens. Devstral Small 2 will cost $0.10 and $0.30, respectively.

For a company, installing the model on its own infrastructure can help keep code private and control customization. For an independent developer or hobbyist, Devstral Small 2 opens the possibility of working locally without sending every file to an external service.

The next thing to watch will be real-world performance beyond the tests: how much time it saves on large projects, how many changes require human review and how much it costs to keep these agents running. Mistral's proposal does not eliminate the programmer, but it does turn the terminal into a space where AI can handle an increasing share of the technical work.