Devstral Small 2 is a 24-billion parameter language model developed by Mistral AI, released in December 2025. It belongs to the Devstral 2 family of models specifically engineered for agentic software development and complex repository-level reasoning. Designed for portability, the model is optimized to run locally on consumer-grade GPUs while providing performance competitive with much larger systems.
The model features a 256,000-token context window, allowing it to maintain architecture-level context across extensive codebases. This capacity supports autonomous workflows such as multi-file orchestration, framework dependency tracking, and automatic retry logic when encountering failures. Unlike the flagship Devstral 2 (123B) model, Devstral Small 2 is released under a permissive Apache 2.0 license, making it suitable for both commercial and open-source applications.
Technically, Devstral Small 2 integrates multimodal capabilities, supporting both text and image inputs for vision-informed coding tasks. It utilizes the Tekken tokenizer with a 131k vocabulary to improve efficiency in code-dense environments. On the SWE-bench Verified benchmark, the model achieved a score of 68.0%, demonstrating strong proficiency in resolving real-world software engineering issues.