Devstral-Small-2505: An Agentic LLM for Software Engineering

Devstral-Small-2505 is a 24 billion parameter language model developed through a collaboration between Mistral AI and All Hands AI. Fine-tuned from Mistral-Small-3.1, this model is specifically engineered for agentic software engineering tasks, demonstrating strong capabilities in tool utilization for codebase exploration and multi-file editing.

Key Capabilities & Features

Agentic Coding: Designed to excel in agentic coding, making it highly suitable for software engineering agents.
High Performance on SWE-Bench: Achieves a verified score of 46.8% on SWE-Bench, outperforming previous open-source state-of-the-art models by 6% and exceeding larger models like Deepseek-V3-0324 and Qwen3 232B-A22B under the OpenHands scaffold.
Lightweight Deployment: With 24 billion parameters, it can run on a single RTX 4090 or a Mac with 32GB RAM, facilitating local and on-device deployment.
Extensive Context Window: Features a 128k token context window, inherited from its base model, Mistral-Small-3.1.
Open-Source License: Released under the Apache 2.0 License, allowing for both commercial and non-commercial use and modification.

Recommended Use Cases

Software Engineering Agents: Ideal for building autonomous agents that can interact with codebases, perform edits, and manage software development workflows.
Code Generation and Refactoring: Its strong performance on SWE-Bench indicates proficiency in generating and modifying code to solve complex problems.
Local Development: Suitable for developers needing a powerful coding assistant that can be run locally on consumer-grade hardware.

Devstral-Small-2505 is optimized for integration with the OpenHands scaffold for optimal performance in agentic workflows.