Mistral AI dropped its next-generation coding models yesterday: Devstral 2 at 123 billion parameters and Devstral Small 2 at 24 billion. The larger model scores 72.2% on SWE-bench Verified, placing it among the top open-weight models for agentic coding tasks. Devstral Small 2 manages 68.0% on the same benchmark while fitting on a single consumer GPU.
Both models ship under permissive licenses (modified MIT for Devstral 2, Apache 2.0 for the smaller version) and support a 256K context window. Mistral is positioning them as dramatically more efficient alternatives to larger competitors: Devstral 2 is 5x smaller than DeepSeek V3.2 and 8x smaller than Kimi K2. The company claims up to 7x better cost efficiency than Claude Sonnet on real-world tasks.
Alongside the models, Mistral introduced Vibe CLI, an open-source terminal assistant that handles file manipulation, code search, and version control through natural language commands. It's already available as a Zed extension.
"Devstral 2 was one of our most successful stealth launches yet, surpassing 17B tokens in the first 24 hours," Kilo Code said. The models integrate with both Kilo Code and Cline out of the box. API access is free during the initial period, with pricing set at $0.40 per million input tokens and $2.00 per million output tokens for Devstral 2 afterward.
The Bottom Line: Mistral's Devstral 2 matches larger models at a fraction of the parameter count, with Devstral Small 2 running locally on GeForce RTX hardware.
QUICK FACTS
- Devstral 2: 123B parameters, 72.2% SWE-bench Verified score
- Devstral Small 2: 24B parameters, 68.0% SWE-bench score, Apache 2.0 license
- Context window: 256K tokens for both models
- API pricing after free period: $0.40/$2.00 per million tokens (Devstral 2), $0.10/$0.30 (Small 2)
- Deployment: Devstral 2 requires minimum 4 H100 GPUs; Small 2 runs on single consumer GPU
