Mistral AI Advances Agentic Coding with Devstral 2 Models and Vibe CLI Release

Advancing Agentic AI in Software Development

Imagine a software engineer sifting through a sprawling repository, tracking dependencies across hundreds of files while an AI agent autonomously suggests fixes and orchestrates multi-file edits in real time. This scenario, once aspirational, is becoming routine as AI models evolve to handle complex, production-grade coding tasks. On December 9, 2025, Mistral AI released Devstral 2, a family of specialized coding models, alongside Mistral Vibe CLI, a terminal-native tool designed to integrate these models into developer workflows. These releases target agentic AI applications, where models act autonomously to explore codebases, detect errors, and implement changes, potentially streamlining software engineering processes amid growing demands for efficiency.

Model Specifications and Benchmarks

Devstral 2 represents a 123 billion parameter dense transformer architecture, equipped with a 256,000 token context window to manage extensive codebases. It achieves 72.2% accuracy on the SWE-bench Verified benchmark, positioning it competitively among open-weight models for software engineering. Released under a modified MIT license, the model is accessible for free through the Mistral API, enabling broad experimentation and deployment. Complementing this is Devstral Small 2, a more compact 24 billion parameter variant sharing the same context length.

Devstral Small 2 Performance

It scores 68.0% on SWE-bench Verified, performing on par with models up to five times larger in scale. Licensed under Apache 2.0, it facilitates production use, including local deployments for privacy-sensitive environments. Both models are optimized for agentic workloads, emphasizing repository-scale operations like dependency tracking, failure detection with retries, and tasks such as bug fixing or legacy system modernization. In comparative evaluations, Devstral 2 demonstrates up to seven times greater cost efficiency than Claude Sonnet 3.5 on real-world coding tasks, a metric critical for continuous agent operations where inference costs accumulate rapidly. Relative to frontier systems, Devstral 2 is five times smaller than DeepSeek V3.2, while Devstral Small 2 is 28 times smaller; against Kimi K2, the reductions are eight times and 41 times, respectively.

These size efficiencies suggest potential for broader accessibility on standard hardware, though real-world performance may vary based on fine-tuning for specific languages or enterprise-scale codebases. Human-led assessments using the Cline agent tool further validate Devstral 2’s edge, showing a 42.8% win rate over DeepSeek V3.2 (versus a 28.6% loss rate) across scaffolded tasks. No direct comparisons to Claude Sonnet 4.5 were detailed in evaluations, but the overall benchmarks indicate parity or superiority in agentic scenarios. Devstral Small 2 extends capabilities to multimodal inputs, processing images alongside code to support agents reasoning over diagrams or screenshots—useful for visual debugging but untested in the provided benchmarks.

Key Performance Metrics:

• SWE-bench Verified: Devstral 2 (72.2%), Devstral Small 2 (68.0%)
• Context Window: 256K tokens for both
• Parameter Counts: 123B (Devstral 2), 24B (Devstral Small 2)
• Cost Efficiency: Up to 7x vs. Claude Sonnet 3.5 on agentic tasks

Tool Integration and Developer Workflow Enhancements

Mistral Vibe CLI, an open-source Python-based command-line interface, operationalizes the Devstral models by enabling natural language interactions directly in terminals or compatible IDEs like Zed, which supports the Agent Communication Protocol. Released under Apache 2.0 and hosted on GitHub, it scans project structures and Git status to maintain contextual awareness, reducing the need for manual context switching. The tool’s architecture supports multi-file orchestration, allowing agents to coordinate architecture-level changes across entire codebases, which could shorten pull request cycles by automating routine edits. Configuration occurs via a simple TOML file, accommodating connections to the Mistral API, local models, or remote endpoints. Features include programmatic execution modes, auto-approval toggles for tools, and granular permissions to mitigate risks in sensitive repositories—essential for enterprise adoption.

• Core Vibe CLI Capabilities:
• Project-aware scanning of file structures and Git status
• Smart autocompletion: @ for files, ! for shell commands, / for config changes
• Persistent chat history with themes optimized for terminal use
• Support for failure retries and multi-step reasoning over code and visuals (via Devstral Small 2)