Devstral 2 - Mistral AI's open-source next-generation programming model series

What is Devstral 2?

Devstral 2 is Mistral AI's next-generation open-source programming model series, including Devstral 2 with 123B parameters and Devstral Small 2 with 24B parameters. Devstral 2 boasts a 256K context window, achieves a SWE-bench validation accuracy of up to 72.2%, and offers superior performance at a cost-effective price. It supports cross-file exploration, architecture-level inference, and autocorrect, making it suitable for enterprise-level development. Devstral Small 2 is lightweight and can be deployed locally on consumer hardware, supporting multimodal input. Both models provide interactive programming assistance via the Mistral Vibe CLI, helping developers complete tasks efficiently and driving the development of open-source programming models.

Main features of Devstral 2

• Cross-file code understanding and generationDevstral 2 can understand and generate code across multiple files, supporting complex code modifications and refactoring in large codebases while maintaining architecture-level context awareness, helping developers efficiently handle multi-file projects.
• Architectural inferenceThe model can track the overall architecture of the codebase, understand framework dependencies, maintain the stability of the overall architecture when making code modifications, and avoid global problems caused by local modifications.
• Automatic error detection and correctionDevstral 2 can detect errors in code, attempt to automatically correct them, and ensure the correctness of the code through multiple trials and verifications, performing exceptionally well in fixing problems in legacy systems.
• Supports multiple programming languagesThe model has good support for multiple mainstream programming languages and can generate and optimize code for different languages according to requirements, meeting diverse development needs.
• Efficient code generation and optimizationDevstral 2 can generate high-quality, efficient code while optimizing the performance of existing code, helping developers improve development efficiency and code quality.

Devstral 2 test performance

• SWE-bench Verified Test
  • Devstral 2 (123B parameters)Accuracy reached 72.2%It has become one of the best-performing open-source models to date.
  • Devstral Small 2 (24B parameters)The accuracy rate is 68.0%Its performance and size are comparable to a model that is five times larger.
• Comparison with competitors
  • Devstral 2 vs. DeepSeek V3.2With a win rate of 42.8%, it is significantly better than DeepSeek V3.2 (28.6%).
  • When compared with Claude Sonnet 4.5 (closed-source model)The continued popularity of Claude Sonnet 4.5 indicates that Devstral 2 still lags behind top-tier closed-source models, but is very close.

Devstral 2 project address

• Project official websitehttps://mistral.ai/news/devstral-2-vibe-cli
• HuggingFace model libraryhttps://huggingface.co/collections/mistralai/devstral-2

Devstral 2 product pricing

• Devstral 2 (123B parameters):It is currently offered free of charge via API.After the free period ends, the API pricing will be [price] per million inputs/output tokens. $0.40 and $2.00.
• Devstral Small 2 (24B parameters):The API is priced at [amount] per million input/output tokens. $0.10 and $0.30.

Application scenarios of Devstral 2

• Code generationDevstral 2 can quickly generate code in multiple languages, helping developers start projects efficiently.
• Code optimizationThe model supports cross-file modifications, helping developers optimize code and improve performance and maintainability.
• Bug FixesThe model automatically detects and corrects code errors, saving debugging time and making it suitable for legacy system repair.
• Enterprise applicationsThe model can be customized and fine-tuned, supports architecture-level inference, and can meet the needs of complex enterprise projects.
• Local deploymentDevstral Small 2 can run locally, protecting data privacy and is suitable for scenarios with high security requirements.