Mistral AI Launches Devstral 2507: Powerful Language Models Tailored for Code

Overview of Devstral 2507 Models

Mistral AI, together with All Hands AI, has introduced updated large language models aimed specifically at developers under the Devstral 2507 series. This release features two models: Devstral Small 1.1 and Devstral Medium 2507. Both are engineered for agent-driven code reasoning, program synthesis, and structured task execution within extensive software repositories. These models emphasize optimized performance and cost-efficiency, making them practical for integration into developer tools and code automation solutions.

Devstral Small 1.1: Open and Efficient

Devstral Small 1.1, also known as devstral-small-2507, is built on the Mistral-Small-3.1 foundation model and includes roughly 24 billion parameters. It supports a vast 128k token context window, enabling it to process multi-file code inputs and lengthy prompts common in software engineering workflows. The model is fine-tuned for structured outputs such as XML and function-calling formats, allowing compatibility with agent frameworks like OpenHands. This makes it suitable for tasks including program navigation, multi-step edits, and code search. Licensed under Apache 2.0, Devstral Small 1.1 is available for both research and commercial use.

Performance on SWE-Bench

Devstral Small 1.1 achieved a 53.6% score on the SWE-Bench Verified benchmark, assessing its ability to generate correct patches for real GitHub issues. This marks a significant improvement over its prior version (1.0) and outperforms other similarly sized open models. The evaluation utilized the OpenHands scaffold, which standardizes testing environments for code agents. While it does not match the largest proprietary models, it strikes a practical balance between size, inference cost, and reasoning capability.

Deployment and Local Inference Options

The model is offered in multiple formats, including quantized GGUF versions compatible with llama.cpp, vLLM, and LM Studio. These allow running inference locally on high-memory GPUs such as the RTX 4090 or Apple Silicon machines with at least 32GB RAM. This flexibility benefits developers and teams seeking independence from hosted APIs. Additionally, Mistral provides API access for the model, priced at $0.10 per million input tokens and $0.30 per million output tokens, consistent with other Mistral-Small models.

Devstral Medium 2507: Higher Performance via API

Devstral Medium 2507 is not open-sourced and is accessible exclusively through the Mistral API or enterprise agreements. It shares the 128k token context length with the Small version but delivers superior performance, scoring 61.6% on SWE-Bench Verified. This places it ahead of commercial competitors such as Gemini 2.5 Pro and GPT-4.1. Its enhanced reasoning abilities over long code contexts suit it for code agents managing large monorepos or cross-file dependencies. API pricing is $0.40 per million input tokens and $2.00 per million output tokens. Enterprise users can also access fine-tuning through Mistral's platform.

Use Cases and Integration

Devstral Small is ideal for local development, experimentation, and embedding in client-side developer tools that require control and cost-efficiency. Conversely, Devstral Medium targets production environments needing higher accuracy and consistency in structured code editing despite higher costs. Both models support integration with code agent frameworks like OpenHands, facilitating structured function calls and XML outputs for automated workflows including test generation, refactoring, and bug fixing. This makes them suitable for IDE plugins, version control bots, and CI/CD pipeline automation.

Summary

The Devstral 2507 release presents a clear choice between cost-effective, open development (Devstral Small) and higher-performance, API-based production usage (Devstral Medium). These models cater to diverse stages of the software engineering lifecycle, from prototype development to enterprise deployment.

For further details, visit the Mistral AI news page.