Mistral AI Releases Devstral 2507 for Code-Centric Language Modeling

Mistral Ai, in cooperation with all Hands Ai, released updated versions of large language models that focus on developers below Devstral 2507 attached. The version includes two models –Devstral Small 1.1 and Devstral average 2507Designed to support thinking about the code based on the agent, synthesis of the program, and implementing organized tasks through large program warehouses. These models are improved for performance and cost, making them useable in the real world in developers tools and code automation systems.

Devstral Small 1.1: An open and integrated model for local and integrated use

Devstral Small 1.1 (It is also called devstral-small-2507It depends on the Mistral-Small-3.1 foundation and contains approximately 24 billion teachers. It supports the window of the distinctive code context 128K, which allows it to deal with the inputs of the multi -file symbol and the typical long demands in the workflow of software engineering.

The form has been set specifically for structured outputs, including XML formats and job formats. This makes it compatible with the agent’s work frameworks such as OpenHAnds and suitable for tasks such as the movement of the program, the editing of multiple steps, and the search for software instructions. It is licensed under APache 2.0 and is available for both research and commercial use.

Performance: SWE engine results

Devstral Small 1.1 achieves 53.6 % On the SWE check -up standard, which evaluates the model’s ability to generate correct corrections to real GitHub problems. This is a noticeable improvement on the previous version (1.0) and puts it before other models available in public with a similar size. The results were obtained using OpenHands, which provide a standard test environment to assess the code factors.

Although it is not at the level of the largest ownership models, this version provides a balance between size, the cost of inference and the performance of practical thinking of many coding tasks.

Publishing: local and good reasoning

The model is released in multiple formats. The quantitative versions are available in GGUF for use with llama.cppand vLLMAnd LM Studio. These formatting formatting locally on high memory graphics processing units (for example, RTX 4090) or Apple’s silicone machines with 32 GB of random access memory or more. This is useful for developers or teams that prefer to work without relying on the hosting of the hosted applications.

Mistral also makes the model available through the inference programming interface. The current pricing is $ 0.10 per million input codes and $ 0.30 per million output symbols, such as other models in the Small Mistral line.

Devstral Medium 2507: higher accuracy, API only

Devstral average 2507 Not open source and is only available through the Mistral Application Programming interface or through the institution’s publishing agreements. It provides the same length of the distinctive code context 128K as a small version but with higher performance.

Model degrees 61.6 % On Swe-Bench, it is scrutinized, outperforming many commercial models, including Gemini 2.5 Pro and GPT-4.1, in the same evaluation framework. Its ability to think of the strongest contexts makes it a candidate for code agents who work via monorabus or large warehouses with crossed dependencies.

API pricing is determined at $ 0.40 per million input codes and $ 2 per million output symbols. The accurate control of the institution’s users is available via the Mistral platform.

Comparison and the use of a suitable condition

Devstral Small is more suitable for local development, experimentation or integration into developers tools by the customer where control and efficiency are important. On the contrary, Devstral Medium provides stronger accuracy and consistency in organized code editing tasks and is intended for production services that benefit from the highest performance despite increased cost.

Integration with tools and agents

Both models are designed to support integration with a code agent framework like OpenHands. Support for structured job calls and XML output formats allows them to have the automatic workflow to generate the test and re -create and install errors. This compatibility makes it easy to connect Devstral models to IDE additions, version control and internal CI/CD pipelines.

For example, developers can use Devstral Small for local workflow models, while Devstral Medium can be used in production services that apply corrections or sorting withdrawal requests based on model suggestions.

conclusion

The Devstral 2507 version reflects an updated update to the LLM staple directed towards the Mistral Code, providing users with a clearer comparison between the cost of reasoning and the accuracy of the task. Devstral Small provides an open -accessed model with sufficient performance for many cases of use, while Devstral Medious meets the right and reliable applications.

The availability of both models in light of the various publishing options makes it relevant through different stages of the workflow of software engineering – from the development of the experimental agent to publishing in commercial environments.

verify Technical detailsand DEVSTRAL Model weights in the face of embrace Devstral Medium will also have a Mistral icon for Institution Customers and on the FineTuning Application interface. All the credit for this research goes to researchers in this project. Also, do not hesitate to follow us twitterAnd YouTube And do not forget to join 100K+ ML Subreddit And subscribe to Our newsletter.

SANA Hassan, consultant coach at Marktechpost and a double -class student in Iit Madras, is excited to apply technology and AI to face challenges in the real world. With great interest in solving practical problems, it brings a new perspective to the intersection of artificial intelligence and real life solutions.