Mistral AI Unveils Devstral 2 Coding Models and Vibe Command-Line Assistant

Details

• Mistral AI has released Devstral 2, a next-generation code-generation large language model available in two versions: the 123 billion parameter Devstral 2 and the smaller 24 billion parameter Devstral Small.
• The main 123 billion model uses a modified MIT license, while the 24 billion model is available under the permissive Apache 2.0 license, allowing wide commercial redistribution.
• Both models can be downloaded or accessed via Mistral’s API without announced usage fees at launch.
• Mistral also debuts Vibe, an Apache 2.0-licensed command-line assistant that can navigate, refactor, and execute code using natural language prompts.
• Vibe installs easily with “uv tool install mistral-vibe” and leverages Devstral for terminal-based developer workflow automation.
• Mistral claims strong accuracy for Devstral on HumanEval, MBPP, and RepoCoder benchmarks, suggesting leading public model performance.
• A technical blog post explains the models were trained on 12 TB of filtered code and language data, multi-stage RLHF, and specialized CLI tool-use fine-tuning.

Impact

The launch of Devstral 2 puts pressure on giants like OpenAI and Google by releasing a high-parameter open source code model and increasing transparency. The dual licensing approach could accelerate enterprise adoption, offering an alternative to restrictive or GPL-licensed options such as Code Llama. With its native CLI agent, Mistral Vibe challenges established developer tools like GitHub Copilot Workspace and positions the company at the forefront of automated software engineering trends.