devstral2

Devstral 2 is a state-of-the-art, 123B-parameter open-source coding model, a dense transformer supporting a 256K context window. It achieves 72.2% on SWE-bench Verified, establishing it as one of the best open-weight models while remaining highly cost-efficient. Released under a modified MIT license, Devstral 2 sets the open state-of-the-art for code agents. Mistral Vibe CLI is a native, open-source command-line coding assistant powered by Devstral, designed to solve software engineering tasks autonomously by exploring, modifying, and executing changes across codebases using natural language.

How to Use

Users can access Devstral 2 via API, which is currently offered for free, or through self-hosted deployment on their own infrastructure with NVIDIA support. Mistral Vibe CLI can be installed directly in the terminal using a curl command (curl -LsSf https://mistral.ai/vibe/install.sh | bash) and then used as a native coding assistant.

Key Features

• State-of-the-art open-source coding model (Devstral 2)
• High performance on SWE-bench Verified (72.2% for Devstral 2, 68.0% for Devstral Small 2)
• Large context window (256K)
• Cost-efficient (up to 7x more than Claude Sonnet)
• Compact and efficient model sizes
• Code Agent Excellence (explores codebases, orchestrates changes, maintains architecture-level context)
• Smart Error Handling (tracks dependencies, detects failures, retries with corrections)
• Multi-File Orchestration across entire codebases
• Native, open-source command-line coding assistant (Mistral Vibe CLI)
• Customizable CLI for specific languages or enterprise codebases
• Project-Aware Context and Smart References in CLI
• Persistent History, autocompletion, and customizable themes in CLI
• IDE Integration for Mistral Vibe CLI (Zed extension, Agent Communication Protocol)
• Image Input Support for Devstral Small 2

Use Cases

• Solving complex software engineering tasks autonomously
• Bug fixing across entire codebases
• Modernizing legacy systems
• Exploring and orchestrating changes in codebases
• Developing AI coding tools and agents
• Powering multimodal agents with image inputs (Devstral Small 2)
• Practical deployment on limited hardware (Devstral Small 2)