Devstral 2: Open-Source 123B Coding Model Hits 72.2% SWE-bench at 1/7th Sonnet's Cost

Mistral AI·high signal

Mistral released Devstral 2 (123B, modified MIT) scoring 72.2% on SWE-bench Verified and Devstral Small 2 (24B, Apache 2.0) scoring 68.0% — both with 256K context. Devstral 2 claims 7x cost efficiency over Claude Sonnet at $0.40/$2.00 per million tokens (currently free via API). Devstral Small 2 at $0.10/$0.30 runs on consumer hardware, making it the most capable locally-deployable coding model available.