# Devstral Devstral is an agentic [[Large Language Models (LLMs)|LLM]] for software engineering tasks, built in collaboration between [[Mistral AI]] and All Hands AI (creators of OpenHands). Released in May 2025 under the [[Apache 2.0 License]], it was later unified into [[Mistral Small 4]]. The model is specifically trained to solve real GitHub issues, running over code agent scaffolds such as OpenHands or SWE-Agent. It scored 46.8% on SWE-Bench Verified, outperforming prior open-source state-of-the-art models by more than 6 percentage points. Under the same test scaffold, it exceeded far larger models like Deepseek-V3-0324 (671B) and Qwen3 232B-A22B, and surpassed GPT-4.1-mini by over 20%. Key characteristics: - **Lightweight**: runs on a single RTX 4090 or Mac with 32GB RAM - **Agentic coding**: trained to contextualize code within large codebases, identify relationships between components, and find subtle bugs - **Open source**: released under [[Apache 2.0 License]] on HuggingFace, Ollama, Kaggle, Unsloth, and LM Studio - **API name**: `devstral-small-2505`, priced same as Mistral Small 3.1 ($0.1/M input, $0.3/M output tokens) - **Enterprise-ready**: suitable for privacy-sensitive repositories with stringent security and compliance requirements ## References - https://mistral.ai/news/devstral - https://huggingface.co/mistralai/Devstral-Small-2505 ## Related - [[Mistral AI]] - [[Large Language Models (LLMs)]] - [[Mistral Small 4]] - [[Apache 2.0 License]]