Mistral AI unveils Devstral 2 family of code models and Mistral Vibe CLI tool

Mistral AI has announced the Devstral 2 model family: the flagship Devstral 2 (123B parameters) with a 256K context window and the compact Devstral Small 2 (24B). Devstral 2 is released under a modified MIT license, while Devstral Small 2 is under Apache 2.0. On the SWE-bench Verified benchmark, the models achieved scores of 72.2% and 68.0%, respectively, placing them alongside the best open models, despite being 5–28 times smaller than DeepSeek V3.2 and 8–41 times smaller than Kimi K2. This provides up to 7x cost savings compared to Claude Sonnet on real-world tasks. Devstral 2 supports codebase exploration, multi-file change orchestration, dependency tracking, and automatic error correction. In human tests with the Cline annotator, Devstral 2 outperforms DeepSeek V3.2 (42.8% wins vs. 28.6% losses) but trails Claude Sonnet 4.5. Devstral 2 is optimized for datacenter GPUs (minimum 4 H100s), while Devstral Small 2 runs on a single GPU or CPU. Both models are fine-tunable. Alongside the models, Mistral Vibe CLI is introduced—an open-source command-line agent with natural language support, file autocompletion, version control, and IDE integration via the Agent Communication Protocol. Currently, Devstral 2 is available for free via API, after which pricing will be $0.40/$2.00 per million tokens (input/output) for Devstral 2 and $0.10/$0.30 for Devstral Small 2.