Mistral AI Unveils Devstral Medium and Upgraded Devstral Small Models

Mistral AI has introduced two new models, Devstral Medium and an upgraded version of Devstral Small, known as Devstral Small 1.1. These models are part of a collaboration with All Hands AI, focusing on enhancing agentic coding capabilities.

Devstral Small 1.1, released under the Apache 2.0 license, maintains its 24 billion parameter architecture but offers significant improvements over its predecessor. It achieves a score of 53.6% on the SWE-Bench Verified benchmark, setting a new standard for open models without test-time scaling. The model is versatile, supporting both Mistral function calling and XML formats, and excels in generalizing to different prompts and coding environments.

Devstral Medium, available through Mistral AI's API, builds on the strengths of Devstral Small and achieves a score of 61.6% on the SWE-Bench Verified benchmark. It offers high performance at a competitive price, making it suitable for businesses and developers seeking cost-effective solutions. The model can be deployed on private infrastructure for enhanced data privacy and supports custom finetuning for specific use cases.

Both models are accessible via API, with Devstral Small 1.1 priced at $0.1 per million input tokens and $0.3 per million output tokens, while Devstral Medium is priced at $0.4 per million input tokens and $2 per million output tokens. These offerings underscore Mistral AI's commitment to providing open-source, high-performance models for the software development community.