Devstral Small 2505 is an agentic language model built specifically for software engineering workflows, developed collaboratively by Mistral AI and All Hands AI. It is fine-tuned from Mistral-Small-3.1 with a custom Tekken tokenizer and stripped of its vision capabilities to focus purely on text-based code tasks. The model is designed to explore codebases, understand multi-file contexts, and make targeted edits across complex repositories. Its architecture supports a generous context window, enabling it to hold entire projects or large code sections in memory during task execution. The combination of its compact 24-billion-parameter footprint and its agentic focus makes it uniquely suited for power users who want strong coding performance without the infrastructure demands of much larger models.

The model's training lineage centers on Mistral-Small-3.1 as its foundation, with additional fine-tuning by All Hands AI to cultivate software engineering expertise. Devstral achieves state-of-the-art results on SWE-Bench Verified at 46.8%, debuting as the top open-source model on that benchmark and outperforming models with significantly more parameters on agentic coding tasks. It integrates natively with the OpenHands scaffold and supports a range of inference frameworks including vLLM, Transformers, and Ollama, with GGUF quantizations available for varied hardware. Fireworks AI notes support for LoRA fine-tuning, allowing organizations to specialize Devstral with their own codebases and conventions. The model is licensed under Apache 2.0 and can run locally on high-end consumer hardware like RTX 4090 GPUs or 32GB RAM Macs, making advanced coding assistance accessible without cloud dependency.