Mistral AI and All Hands AI have jointly released Devstral Medium and an upgraded Devstral Small 1.1, focusing on generalization, performance, and cost-efficiency for code agents.
Mistral AI, in collaboration with All Hands AI, has introduced two significant additions to its AI model suite: Devstral Medium and an upgraded Devstral Small 1.1. These releases emphasize robust generalization across prompt types and compatibility with agentic scaffolds.
Devstral Small 1.1: Open and Improved
Maintaining its 24B parameter architecture, Devstral Small 1.1 brings notable advancements over its predecessor. Released under the Apache 2.0 license, it continues the tradition of open accessibility while offering:
- Enhanced performance: Scores 53.6% on SWE-Bench Verified, setting a new benchmark among open models for code agents.
- Increased versatility: Integrates effectively with OpenHands and supports both Mistral function calling and XML formats, broadening its application range.
Devstral Medium: Advanced Capabilities via API
Devstral Medium extends the capabilities of its smaller counterpart with a 61.6% score on SWE-Bench Verified. Positioned on the cost/performance pareto frontier, it outperforms commercial models like Gemini 2.5 Pro and GPT-4.1 at a fraction of the cost. Key features include:
- API availability for developers and businesses
- On-premise deployment options for data-sensitive applications
- Custom finetuning support, allowing enterprise users to tailor the model to specific needs
Pricing and Access
Both models are available via API under the following configurations:
| Model Name | Price (Input/Output Tokens) |
|---|---|
| devstral-small-2507 | $0.1/M input, $0.3/M output tokens |
| devstral-medium-2507 | $0.4/M input, $2/M output tokens |
Devstral Small 1.1 remains open-source and deployable locally. Interested users can explore its model card on Hugging Face. Devstral Medium is accessible through Mistral Code and the finetuning API, with enterprise-level support available through direct contact.
Mistral AI reiterates its commitment to open-source innovation, ensuring high-quality models are available to the broader development community.