Quick Take: Mistral AI, teaming up with All Hands AI, just introduced Devstral – an agentic LLM specifically designed for complex software engineering tasks. Trained to solve actual GitHub issues and reportedly outperforms all open-source models on SWE-Bench Verified by a significant margin. It’s being released under an Apache 2.0 license, making it accessible for local deployment, enterprise use, and copilots.
Why Devstral is (hopefully) a Big Deal
Devstral is engineered to tackle real GitHub issues and operates using code agent scaffolds like OpenHands or SWE-Agent, which define how the model interacts with test cases.
The performance claims are bold. On the SWE-Bench Verified benchmark (a dataset of 500 real-world GitHub issues screened for correctness), Devstral achieves a score of 46.8%. This isn’t just a small bump; it reportedly outperforms previous open-source state-of-the-art models by more than 6 percentage points.
When evaluated using the same test scaffold (OpenHands, provided by All Hands AI), Devstral even surpasses much larger models like Deepseek-V3-0324 (a whopping 671B parameters) and Qwen3 232B-A22B. Even when compared against closed and open models evaluated under any scaffold, Devstral shows substantially better performance than several closed-source alternatives, for instance, surpassing the recent GPT-4.1-mini by over 20%.
Versatility: From Your Laptop to Enterprise and Beyond
One of Devstral’s key advantages is its versatility. It’s light enough to run on a single RTX 4090 or even a Mac with 32GB RAM, making it a prime candidate for local deployment and on-device use. This means coding platforms like OpenHands can allow Devstral to interact directly with local codebases, offering fast resolution to issues right where developers work. Its strong performance also makes it a suitable choice for agentic coding on privacy-sensitive repositories within enterprises, especially those with stringent security and compliance requirements. And if you’re building or using an agentic coding IDE, plugin, or environment, Devstral is positioned as a great model to add to your selection.
Open and Accessible
In a significant move for the community, Mistral is releasing Devstral for free under an Apache 2.0 license. You can grab the model from their model card, and it’s also available on Mistral’s API under the name devstral-small-2505.
For those who prefer to self-deploy, Devstral can be downloaded from HuggingFace, Ollama, Kaggle, , and LM Studio.
TLDR
Devstral, an open-source AI agent built to solve real GitHub issues, not just write code snippets. Light enough for local use (RTX 4090/Mac), powerful enough for enterprise, and free under Apache 2.0. A bigger version is coming soon.
