French AI startup Mistral releases Devstral Small 2, a fully open source, laptop-friendly coding model, alongside Vibe CLI, enabling offline, private software development.
French AI startup Mistral has unveiled Devstral 2, a high-performance coding model, and Devstral Small 2, a laptop-friendly, fully open source variant designed for offline software development. The release also introduces Vibe CLI, a command-line interface that integrates directly with the models for project-wide code understanding, orchestration, and refactoring.
Devstral 2 is a 123-billion parameter transformer with a 256K-token context window, optimised for agentic software engineering, achieving 72.2% on SWE-bench Verified, and requires 4× H100 GPUs for deployment. Its smaller sibling, Devstral Small 2, packs 24 billion parameters, matches the same long-context performance, scores 68.0% on SWE-bench Verified, and runs efficiently on a single GPU or standard laptop. It is fully licensed under Apache 2.0, allowing unrestricted commercial or open-source use.
The Vibe CLI complements the models with a terminal-native agent capable of understanding file structures, managing dependencies, executing commands, and refactoring across multiple files. Released under Apache 2.0, it provides developers autonomy, privacy, and freedom from SaaS or vendor lock-in.
Mistral’s licensing structure differentiates the models: Devstral Small 2 is fully open, while Devstral 2 uses a modified MIT license restricting use for companies with >$20 million monthly revenue, requiring a commercial license.
Developers have praised the portability and offline usability, with Victor Mustar, Head of Product at Hugging Face, asking: “Is the small, Apache 2.0 licensed variant the new local coding king?”
