Just a few days after announcing its range of AI models Mistral 3, the French Artificial Intelligence company Mistral AI has announced its second family of models designed for developers: Devstral 2. The release includes two models optimized for software engineering tasks: a large one, Devstral 2, and another, Devstral Small 2, small enough to work perfectly even on a conventional laptop and without the need for a connection.
Along with these models, Mistral has announced Vibe, an agent accessible through a command line interface, designed so that developers can access models directly from their terminals.
Devstral 2 is offered for a time for free through the Mistral API and Hugging Face. When this free period ends, the API Price for Devstral 2 will be $0.40 per million tokens in, and $2 per million tokens out. For Devstral Small 2 it will be $0.10 per million input tokens, and $0.30 per million output tokens.
The larger of the two models, Devstral 2, has 123 billion parameters and a context window of 256,000 tokens, and is built specifically for agentic software development. For its part, Devstral Small 2, with Apache 2.0 license, has 24 billion parameters, with the same long context window as the family’s LLM. It is intended for internal use by teams in companies, local deployments and low-latency edge inference. Due to its characteristics, it is indicated for companies in heavily regulated environments and companies that need to run AI models within their facilities and infrastructure.
According to Mistral, Devstral 2, with a modified MIT license (implies that companies that want to use it and enter more than 20 million monthly will have to accept a commercial license) achieves 72.2% effectiveness in the SWE-bench Verified test bench, designed to evaluate broad-context software engineering tasks in real repositories. Devstral Small 2 reaches 68% on it. This is the most effective result for open weight models of its size.
Mistral Vibe, Mistral’s command-line assistant, integrates directly with Devstral models. It is a native interface designed for understanding and orchestrating project code, which is incorporated into the developer’s workflow.
Vibe is capable of, among other things, reading the Git file and status tree to understand the scope of a projectand allows you to reference files using an @, as well as execute commands with !. Vice also orchestrates changes to multiple files, tracks dependencies, resumes failed runs, and can do architecture-scale refactoring. It is programmable, and also has the Apache 2.0 license, which means that its use is free, even in commercial deployments or internal tools.
