Mistral AI Launches Devstral 2 and Vibe CLI
- •Mistral AI has released the Devstral 2 model family, featuring 123B and 24B parameter versions optimized for software development.
- •The flagship 123B model achieved a 72.2% success rate on the SWE-bench Verified benchmark, surpassing larger competitors like DeepSeek V3.2.
- •A new open-source Vibe CLI enables autonomous AI agents to manage, edit, and update complex codebases directly from the terminal.
Mistral AI has launched Devstral 2, a family of large language models engineered for advanced programming tasks. The flagship 123B parameter model features a massive 256K context window, enabling it to process extensive code repositories in a single pass. It achieved a 72.2% success rate on the SWE-bench Verified benchmark, outperforming several larger competitors as a premier open-source option for coding. This foundation model is designed to handle complex logic and long-range dependencies in professional software engineering.
Mistral also introduced Devstral Small 2, a compact 24B version optimized for efficiency on standard consumer hardware. This model allows developers to run powerful assistants locally, ensuring data privacy and reducing reliance on cloud infrastructure. By lowering the hardware barrier, Mistral enables creators to leverage state-of-the-art AI for localized development tasks. Both models support fine-tuning, a process that adapts the AI to proprietary languages or specific enterprise workflows, making them highly versatile for diverse technical environments.
Complementing the models is Vibe CLI, an open-source terminal tool acting as an autonomous AI agent. This tool explores codebases and executes file edits independently to resolve bugs or modernize legacy systems. By integrating into the terminal, it streamlines the software lifecycle through automated reasoning and direct file manipulation. Evaluations suggest Devstral 2 is more cost-efficient than closed-source alternatives like Claude 3.5 Sonnet, providing a robust, transparent solution for the global software development ecosystem.