Now loading...
Mistral AI has unveiled its latest advancements in AI-driven coding with the launch of Devstral 2 and Devstral Small 2, a pair of powerful open-source models designed to push the boundaries of software engineering automation. The larger Devstral 2 boasts 123 billion parameters, while the smaller variant packs 24 billion, both aimed at making high-performance code generation accessible to a wider audience through permissive licensing. Devstral 2 operates under a tweaked MIT license, and Devstral Small 2 follows the Apache 2.0 standard, fostering broader adoption in distributed AI efforts.
Right now, developers can access Devstral 2 at no cost through Mistral’s API, with plans for paid tiers down the line. Alongside the models, the company introduced Mistral Vibe, a command-line interface tailored for Devstral that streamlines full-cycle code tasks directly from the terminal.
Standing out in benchmarks, Devstral 2 achieves a leading 72.2 percent score on the SWE-bench Verified test, marking it as a top contender among open-weight models for code agents despite using far fewer resources than rivals. It delivers up to seven times the cost efficiency of Anthropic’s Claude Sonnet in practical scenarios. The smaller Devstral Small 2 holds its own with a 68 percent performance on the same benchmark, competing effectively against systems several times its scale and running smoothly on everyday hardware.
These models represent a leap in efficiency, with Devstral 2 being five times slimmer than DeepSeek V3.2 and eight times more compact than Kimi K2, while Devstral Small 2 shrinks those gaps even further at 28 and 41 times smaller, respectively. This compactness opens doors for deployment on modest setups, benefiting independent coders, startups, and enthusiasts who might otherwise struggle with resource-heavy alternatives.
Tailored for real-world development pipelines, Devstral 2 excels at navigating entire code repositories, coordinating edits across files, and preserving high-level design insights. It handles dependencies between frameworks, spots errors, and iterates fixes autonomously, tackling issues from debugging to updating outdated codebases. Developers can further adapt it via fine-tuning for particular programming languages or massive corporate repositories.
In head-to-head assessments by an external evaluator using the Cline framework, Devstral 2 outperformed DeepSeek V3.2 with a 42.8 percent win rate against a 28.6 percent loss rate. Still, it trails behind premium closed models like Claude Sonnet 4.5, highlighting ongoing challenges for open alternatives in matching proprietary precision. As Cline noted, Devstral 2 stands at the cutting edge of open coding tools, offering tool-calling reliability that rivals the leaders and providing a seamless experience for users. Kilo Code echoed this, calling the release a stealth success that racked up over 17 billion tokens in its first day, praising Mistral’s rapid pace in delivering scalable, budget-friendly solutions.
Devstral Small 2 mirrors the flagship’s 256,000-token context length and adds versatility with support for image processing, enabling multimodal applications in agentic workflows. Its lightweight design ensures quick responses, seamless iteration, and on-device privacy without cloud dependency.
Mistral Vibe CLI brings these capabilities to the developer’s desktop as an open-source tool under Apache 2.0, allowing natural language instructions to probe, alter, and run code changes across projects. It integrates into terminals or IDEs through the Agent Communication Protocol, featuring an interactive chat mode equipped for file handling, searches, Git operations, and shell commands. The tool auto-detects project layouts and version states for contextual awareness, offers handy shortcuts like @ for files and ! for commands, and reasons across full codebases to speed up pull request reviews by as much as half. Additional perks include conversation memory, smart completions, and theme options, with easy scripting, approval toggles, provider setups via a config file, and permission controls.
To dive in, Devstral 2 remains free on the API for now, shifting to $0.40 per million input tokens and $2.00 for output later, while Devstral Small 2 will cost $0.10 and $0.30 respectively. Mistral has teamed up with agent platforms like Kilo Code and Cline for seamless integration into existing workflows. Vibe CLI also plugs into the Zed editor as an extension.
For setup, Devstral 2 suits data center environments with at least four H100-level GPUs and is testable on NVIDIA’s build platform. Devstral Small 2 thrives on a single GPU across NVIDIA lines like DGX Spark or RTX cards, with CPU-only options viable and NVIDIA NIM compatibility incoming. Aim for a generation temperature of 0.2 and check the Vibe CLI guidelines for peak results.
Mistral invites feedback and shares on X, Discord, or GitHub, and is actively recruiting talent to advance open AI frontiers at its careers page.
