French startup Mistral AI has released Devstral 2, a 123 billion parameter open-weights AI model for coding, scoring 72.2 percent on the SWE-bench Verified benchmark. Alongside it, the company introduced Mistral Vibe, a command-line interface for autonomous software engineering tasks. A smaller version, Devstral Small 2, also debuted for local use on consumer hardware.
On December 10, 2025, Mistral AI unveiled Devstral 2, designed to function within an autonomous software engineering agent. This model excels at resolving real GitHub issues, achieving a 72.2 percent score on SWE-bench Verified, a test involving 500 problems from popular Python repositories. The benchmark requires the AI to read issue descriptions, navigate codebases, and produce patches that pass unit tests—tasks often seen as straightforward bug fixes by experienced engineers.
Complementing the model is Mistral Vibe, a CLI tool licensed under Apache 2.0. It enables developers to interact with Devstral models directly in their terminal, scanning file structures and Git status for project-wide context. The tool can modify multiple files and run shell commands independently, akin to interfaces like Claude Code or OpenAI Codex.
Mistral also launched Devstral Small 2, a 24 billion parameter variant scoring 68 percent on the benchmark. It operates offline on laptops and both models handle a 256,000 token context window for sizable codebases. Devstral 2 uses a modified MIT license, while the smaller one is under Apache 2.0.
Pricing starts free via Mistral's API, shifting to $0.40 per million input tokens and $2.00 per million output tokens for Devstral 2—claimed to be seven times more efficient than Anthropic's Claude Sonnet 4.5, which charges $3 and $15 per million tokens respectively.
The release ties into 'vibe coding,' a term coined by Andrej Karpathy in February 2025, describing natural language prompts for AI-generated code without deep review. Developer Simon Willison praised it for prototyping: “I really enjoy vibe coding. It’s a fun way to try out an idea and prove if it can work.” Yet he cautioned, “vibe coding your way to a production codebase is clearly risky,” emphasizing the need for code quality in evolving systems.
Mistral asserts Devstral 2 can sustain project coherence, fix bugs, modernize legacy code, and manage dependencies at scale, potentially extending vibe coding beyond prototypes.
