Upgrading agentic coding capabilities with the new Devstral models

This Research article from Mistral AI introduces a new round of upgrades to the Devstral series, focusing on "agentic coding" capabilities, meaning the model can understand tasks, modify code, handle environment feedback, and advance software development work within a code agent framework. This update includes two models: the open-source-oriented Devstral Small 1.1, and the higher-performance Devstral Medium, aimed at API and enterprise deployment. Devstral Small 1.1 continues the previous version's 24B-parameter architecture and is released under the Apache 2.0 license; Mistral emphasizes that it scores 53.6% on SWE-Bench Verified and leads among open models that do not use test-time scaling. It is also described as generalizing better across different prompts, different coding environments, and different agentic scaffolds, and it supports Mistral function calling and XML format, making it easy to integrate with tools such as OpenHands. Devstral Medium further pushes SWE-Bench Verified up to 61.6%; the company claims it offers a new trade-off point between cost and performance and can be used through Mistral's public API. On pricing, devstral-small-2507 costs $0.1 per million input tokens and $0.3 per million output tokens; devstral-medium-2507 costs $0.4 for input and $2 for output. For enterprises, the focus of Devstral Medium is not only the API but also private infrastructure deployment, data control, and fine-tuning for specific scenarios through the finetuning API. Overall, this is a productization-leaning research release, of greater reference value to developers, ML engineers, and teams that need to build coding agents.