Mistral releases Devstral 2 and Devstral Small 2, plus the Vibe CLI
Mistral has published a new pair of coding models and an open-source terminal agent intended for production code workflows. The lineup includes **Devstral 2 (123B parameters, modified MIT license)** and **Devstral Small 2 (24B parameters, Apache 2.0)**, both supporting a **256K token context window** and designed for on-prem deployment and fine-tuning.
Models and performance
Devstral 2 is positioned as an open-weight model optimized for code agents, achieving **72.2% on SWE-bench Verified**. Devstral Small 2 posts **68.0% on SWE-bench Verified** while remaining compact enough to run on consumer hardware. Mistral reports that both models are substantially smaller than several competitors—Devstral 2 and Small 2 are respectively 5x and 28x smaller than DeepSeek V3.2, and 8x and 41x smaller than Kimi K2—emphasizing model compactness alongside performance.
In human evaluations conducted through an independent annotation provider, Devstral 2 demonstrated a **42.8% win rate** against DeepSeek V3.2 (with a 28.6% loss rate), while still lagging behind Claude Sonnet 4.5 in preference. The announcement also cites **up to 7× better cost efficiency** than Claude Sonnet on real-world tasks.
Built for multi-file, production workflows
The models are described as capable of exploring entire codebases and orchestrating changes across multiple files, tracking dependencies, detecting failures, and retrying with corrections. Fine-tuning options exist to emphasize particular languages or adapt to large enterprise codebases. Devstral Small 2 additionally supports image inputs for multimodal agents and is suitable for locally deployed, private inference.
Mistral Vibe CLI — terminal-native agent
Mistral Vibe CLI is an open-source, Apache-licensed command-line coding assistant powered by Devstral. It integrates into the terminal or an IDE via the Agent Communication Protocol and provides tools for file manipulation, code search, version control, and command execution. Key capabilities include:
- **Project-aware context** that scans file structure and Git status
- **Multi-file orchestration** for architecture-level reasoning across a repository
- Smart in-terminal references and commands (file autocomplete, shell execution, slash commands)
- Persistent history, autocompletion, and configurable themes
Vibe CLI can run programmatically, supports toggling auto-approval for tool execution, and allows configuration of local models and providers through a config.toml. The Vibe CLI extension is available for Zed: https://zed.dev/extensions. Best-practice guidance is linked in the project README: https://github.com/mistralai/mistral-vibe/blob/main/README.md.
Availability, pricing, and deployment
Devstral 2 is currently free to use via Mistral’s API at https://console.mistral.ai/. After the free period, pricing will be **$0.40/$2.00 per million tokens (input/output) for Devstral 2** and **$0.10/$0.30 for Devstral Small 2**.
Deployment guidance:
- **Devstral 2**: optimized for data-center GPUs; minimum **4 H100-class GPUs** recommended. Try it on http://build.nvidia.com.
- **Devstral Small 2**: designed for single-GPU operation and runs across NVIDIA systems and consumer GPUs; also supports CPU-only configurations. NVIDIA NIM support is planned.
Mistral notes partnerships with agent tool providers Kilo Code (https://kilo.ai/) and Cline (https://cline.bot/) to integrate Devstral into existing toolchains.
Community and hiring
Community touchpoints and contribution channels include X/Twitter (https://x.com/mistralai), Discord (https://discord.com/invite/mistralai), and GitHub (https://github.com/mistralai). Open positions and career information are on the Mistral careers page: https://mistral.ai/careers.
Original source: https://mistral.ai/news/devstral-2-vibe-cli
