Devstral 2 and Mistral Vibe CLI (mistral.ai/news/devstral-2-vibe-cli/, released by Mistral AI) introduce a next-generation open coding ecosystem comprising frontier coding foundation models and a native terminal agent. Shipped in two sizes—Devstral 2 (123B parameters, modified MIT license) and Devstral Small 2 (24B parameters with vision support, Apache 2.0)—the models deliver state-of-the-art software engineering reasoning, achieving 72.2% on SWE-bench Verified at up to 7× greater cost efficiency than closed competitors. Alongside the models, Mistral open-sourced Mistral Vibe CLI (Apache 2.0), an interactive terminal agent supporting project-aware file manipulation, shell command execution, Agent Client Protocol (ACP), and Model Context Protocol (MCP) integrations.

As software engineering shifts from inline autocomplete to autonomous agents, developers require models capable of navigating entire multi-file repositories, diagnosing build failures, and applying test-driven corrections. Devstral 2 is explicitly architected for agentic workflows, featuring a 256K context window to ingest entire codebases and outperforming significantly larger models while maintaining a compact parameter footprint. To ground these models locally, Mistral Vibe CLI brings autonomous code planning and execution directly into the developer’s command line.

Core Architecture & Technical Profile at a Glance

Feature / Dimension	Devstral 2 (123B) & Devstral Small 2 (24B)	Mistral Vibe CLI Specifications
Primary Focus	Autonomous code agents, bug fixing, refactoring	Terminal-native autonomous software engineering agent
Model Footprint	123B (Dense) & 24B (Consumer GPU / Mac runnable)	Lightweight Python-based CLI utility
Context Window	256,144 tokens (256K long-context)	Scans directory structures and Git status automatically
Benchmark Records	72.2% on SWE-bench Verified (Devstral 2)	42.8% human win rate vs. DeepSeek V3.2 in Cline evaluations
Vision & Multimodal	Devstral Small 2 supports UI/diagram-to-code	Browser rendering and breakpoint verification loops
Extensibility & Standards	Native function/tool calling for agentic loops	Agent Client Protocol (ACP) & Model Context Protocol (MCP)
Licensing	Modified MIT (123B) / Apache-2.0 (24B)	Apache-2.0 Open Source
API Pricing	$0.40/M input, $2.00/M output (Devstral 2)	Free CLI tool; connects via API key or local models

Key Innovations and Developer Capabilities

• High Performance with Compact Efficiency: At 123B parameters, Devstral 2 is 5× smaller than DeepSeek V3.2 and 8× smaller than Kimi K2, yet it matches or exceeds their SWE-bench performance while running at substantially lower latency and cost.

• Devstral Small 2 (Edge & Local Execution): The 24B variant scores 68.0% on SWE-bench Verified and runs on a single consumer GPU (RTX 4090) or unified-memory Mac (32GB RAM). It adds vision support for interpreting UI mockups, architectural diagrams, and error screenshots directly into code.

• Mistral Vibe CLI Workflow Features:

  • Smart Syntax & Interactivity: Reference project files using @ autocompletion, execute bash commands with !, and adjust system behaviors using slash commands.

  • Agent Client Protocol (ACP) & MCP Support: Conforms to ACP standards, enabling Vibe to drive external IDEs (like VS Code or JetBrains) and connect to custom MCP servers (e.g., live Figma design tokens, GitHub issues).

  • Autonomous Failure Recovery: Detects compilation errors, runs test suites, and iteratively rewrites problematic files until linting and integration tests pass.

• Enterprise Integration and Ecosystem Support: Integrated out-of-the-box with leading open-source agent scaffolds including Cline and Kilo Code, allowing teams to swap proprietary coding APIs for open-weights alternatives with zero friction.