Mistral AI has introduced a large language model, Devstral, optimized for addressing challenges that arise during software development. Unlike standard AI models, Mistral AI goes beyond simply writing individual functions and code suggestions, providing capabilities to analyze and contextualize (determine the purpose and logic) large codebases, identify relationships between components, and pinpoint hard-to-detect bugs in complex functions.
The model encompasses 23.6 billion parameters, considers a context of 128 thousand tokens, and is released under the Apache 2.0 license. The downloadable archive with Devstral is 47 GB and suitable for use on local systems — a single PC with an NVIDIA GeForce RTX 4090 graphics card and 32 GB of RAM is sufficient to run the model. The model can be utilized in open toolkits like SWE-agent and OpenHands for automating bug fixes, code analysis, and modifications.
The system can be employed to resolve specific issues on GitHub and significantly outperforms other projects in the SWE-Bench Verified test suite, which evaluates the accuracy of solutions for typical code problems (offering 500 tests based on real error reports on GitHub). In this test, the Devstral model scored 46.8%, while the Claude 3.5 Haiku model achieved 40.6%, SWE-smith-LM 32B scored 40.2%, and GPT-4.1-mini scored 23.6%. Among others, Devstral outperformed major models such as Deepseek-V3-0324 671B (38.8%) and Qwen3 232B-A22B (34.4%), which encompass hundreds of billions of parameters.
Source: opennet.ru
