Mistral ai introduced a large language model Devstral , optimized for solving problems arising in the process of software development. Unlike typical AI models, Mistral AI goes beyond the scope of individual functions and additions of code, and provides opportunities that allow analyzing and contextualize (determine the purpose and logic of work) large codes, determine the connections between components and determine the difficult to make errors in confusing functions.
The model covers 23.6 billion parameters, takes into account the context of 128 thousand tokens and published under the Apache 2.0 license. The loading archive with Devstral occupies 47 GB and is suitable for use on local systems-there is enough one PC with the NVIDIA GeForce video card. RTX 4090 and 32 GB of RAM. The model can be used in the tools swe-gent and openhands for automation of error correction, code analysis and amendments.
Devstral model is trained to solve specific problems on GITHUB and noticeably ahead of other systems in the Swe-Bench Verified test set that checks the correctness of the standard problems in the code based on 500 real errors on GitHub. In this test, the Devstral model scored 46.8%, while the Claude 3.5 Haiku model received 40.6%, SWE-SMITH-LM 32B-40.2%and GPT-4.1-Mini-23.6%. Among other things, Devstral was ahead of such large models as Deepseek-V3-0324 671b (38.8%) and QWEN3 232B-A22B (34.4%), covering hundreds of billions of parameters.