Introducing Devstral 2 and Mistral Vibe CLI

Mistral AI has launched Devstral 2 and the Mistral Vibe CLI, with this release focusing on open-source code agents and real-world software engineering workflows. Devstral 2 is a 123B-parameter dense transformer supporting a 256K context window, and the company claims it achieves a 72.2% score on SWE-bench Verified, positioning it as a high-performance open-weight coding model; Devstral Small 2 is the 24B version, also supporting a 256K context, scoring 68.0% on SWE-bench Verified, and is designed to be deployable on smaller hardware, a single GPU, or even some consumer-grade or CPU-only environments. In terms of licensing, Devstral 2 uses a modified MIT license and Devstral Small 2 uses Apache 2.0, and the company emphasizes that both are open-source with relatively permissive licenses. Mistral also makes cost efficiency a selling point, stating that Devstral 2 can be more cost-efficient than Claude Sonnet on real-world tasks, while at the same time acknowledging that Claude Sonnet 4.5 still leads clearly in human preference evaluations, indicating that a gap remains between open-source and closed models. On the tooling side, the Mistral Vibe CLI is an open-source command-line coding assistant paired with Devstral that can scan project structure and Git status, supports natural language interaction, file operations, code search, version control, and command execution, and can also reference files using @, run shell commands using !, and adjust settings via slash commands. It can be used in the terminal, and can also be integrated into IDEs via the Agent Communication Protocol, while offering options such as auto-approval, tool permissions, and local model and provider configuration. Overall, this is a comprehensive move by Mistral into the coding agent market: on one hand it launches the self-hostable and fine-tunable Devstral model series, and on the other it provides the Vibe CLI as a practical development entry point, integrating with tool ecosystems such as Cline, Kilo Code, and Zed. For Taiwanese developers and ML engineering teams, the noteworthy aspects are the open-source license, long context, automatic multi-file modification, local deployment, and the privacy of internal enterprise code; but actual adoption still requires independent verification of the model's stability on existing repos, frameworks, testing workflows, and Traditional Chinese requirements.