Mistral releases Devstral 2 and Vibe CLI
Mistral reported the 123B Devstral 2 scoring 72.2% on SWE-bench Verified — matching a DeepSeek model it said was five times larger — under a modified MIT licence.
- Models & capabilities
- Open weights & ecosystem
- Minor
Mistral released Devstral 2, a pair of open-weight coding models, alongside Mistral Vibe, a command-line coding agent built to work with them. Devstral 2 has 123 billion parameters and a 256,000-token context window, released under a modified MIT licence; a smaller Devstral Small 2, at 24 billion parameters and additionally able to take image input, was released under Apache 2.0.
Mistral reported Devstral 2 scoring 72.2% on SWE-bench Verified, a benchmark of real GitHub issue resolution, against 68.0% for the smaller model. The company framed both as competitive with DeepSeek’s V3.2 despite being, by its own account, five and twenty-eight times smaller respectively, and cited human evaluations in which Devstral 2 was preferred over DeepSeek V3.2 in 42.8% of comparisons against a 28.6% loss rate — while noting Anthropic’s Claude Sonnet 4.5 remained more preferred than either. Mistral Vibe adds project-aware context scanning, multi-file editing and shell execution on top of the models.
Both models were made available free through Mistral’s API at launch, with the company stating a future price of $0.40 per million input tokens and $2.00 per million output tokens. The release continued the pattern of open-weight coding models from European and Chinese labs being positioned as capable substitutes for proprietary tools at a fraction of the parameter count, part of the broader argument over whether frontier-level coding performance requires frontier-scale models.