The French AI startup Mistral has launched Devstral Small 24B, a new open language model built for software development and described as "agentic."
Developed in partnership with All Hands AI, Devstral is released under the Apache 2.0 license and sets a new record for open models on the SWE-Bench Verified benchmark, scoring 46.8 percent. The SWE-Bench Verified benchmark uses a dataset of 500 real, manually verified GitHub issues.
That result puts Devstral ahead of both Deepseek-V3 and Google's Gemma 3, according to Mistral. Devstral replaces Mistral's previous Codestral model, which was not available for commercial use. Mistral says Devstral is designed for integration into agentic coding IDEs, plugins, and development environments.
A larger version of the model is expected in the coming weeks.
Compared to proprietary models, Devstral also performs significantly better, outperforming the recently released GPT-4.1-mini by over 20 percent, according to the company. Mistral notes that Devstral is a "research preview" and feedback is welcome.
Devstral is optimized to run locally on systems with an RTX 4090 GPU or a Mac with 32 GB of RAM, and supports tools like OpenHands and SWE-Agent. The model is available on Hugging Face, Ollama, Kaggle, LM Studio, and through Mistral's own API. Pricing is set at $0.10 per million input tokens.
