AI BriefingStrengthening Agentic Coding Capabilities with New Devstral Models
Key point
Mistral AI has released Devstral Small 1.1 and Devstral Medium models with significantly enhanced coding agent performance.
Details
The Devstral model series, born from a collaboration between Mistral AI and All Hands AI, focuses on improving generalization across diverse prompts and agent scaffolds.
Devstral Small 1.1 is an open model with 24B parameters released under the Apache 2.0 license. It scored 53.6% on SWE-Bench Verified, achieving top-tier performance among open models without test-time scaling. Notably, it supports both Mistral function calling and XML formats, demonstrating excellent generalization in environments like OpenHands.
Devstral Medium is available via API and recorded a high score of 61.6% on SWE-Bench Verified. It offers excellent cost-efficiency at about a quarter of the price of Gemini 2.5 Pro and GPT 4.1, and also supports on-premise deployment and custom fine-tuning for enterprises.
Detailed pricing by model is as follows.
- devstral-small-2507: input $0.1/M, output $0.3/M (same as Mistral Small 3.1)
- devstral-medium-2507: input $0.4/M, output $2/M (same as Mistral Medium 3)
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.