Devstral 2
2025-12-09 · Publisher : Mistral AI · Domain : Artificial Intelligence
Mistral AI launches Devstral 2, a dense 123-billion-parameter model for coding agents, at 72.2% on SWE-bench Verified, under a modified MIT license, with the Mistral Vibe command-line interface.
What changes
- 72.2% on SWE-bench Verified; announced as the best open model for coding agents with far fewer parameters than its competitors.
- Up to 7 times more economical than Claude Sonnet on real tasks according to the vendor; Claude Sonnet 4.5 remains clearly preferred in human evaluation.
- Launch of Mistral Vibe CLI, an open source terminal agent (Apache 2.0).
- Free via API during the launch period.
### Specifications - 123 billion parameters, dense transformer, 256k context; text. - License: modified MIT (open weights). - Deployment: 4 H100-class GPUs minimum. - Availability: API, Hugging Face, build.nvidia.com; Kilo Code, Cline, Zed. - API identifier: `devstral-2512` (deprecated on May 22, 2026 according to the documentation). ### Pricing | Model | Input (USD / M tokens) | Output (USD / M tokens) | |---|---|---| | Devstral 2 (after the free period) | 0.4 | 2 | Free via the API during launch; the documentation indicates a switch to paid in January 2026. ### Benchmarks Scores published by the vendor (Hugging Face model card): | Benchmark | Devstral 2 | Devstral Small 2 | |---|---|---| | SWE-bench Verified | 72.2% | 68.0% | | SWE-bench Multilingual | 61.3% | 55.7% | | Terminal Bench 2 | 32.6% | 22.5% | Human evaluation (Cline scaffold): Devstral 2 wins against DeepSeek V3.2 at 42.8% versus 28.6% losses. ### Key takeaways A leading open coding agent model, with an associated terminal tool, but one that remains behind the most capable closed models.
Links
Tags : LLM, Mistral, Devstral, Code, Agents, Open Source