What is Devstral 2
Devstral 2 is a new generation of open-sourcecodingmodel series launched by Mistral AI, including the 123B-parameter Devstral 2 and the 24B-parameter Devstral Small 2. Devstral 2 has a 256K context window and achieves a SWE-bench verified accuracy of up to 72.2%, delivering outstanding performance and cost efficiency. It supports features such as cross-file exploration, architecture-level reasoning, and automatic correction, making it suitable for enterprise-level development. Devstral Small 2 is lightweight, can be deployed locally on consumer-grade hardware, and supports multimodal input. Both provide terminal-based interactive coding assistance through the Mistral Vibe CLI, helping developers complete tasks efficiently and advancing the development of open-source coding models.
Main Features of Devstral 2
- Cross-file code understanding and generation: Devstral 2 can understand and generate code across multiple files, support complex code modifications and refactoring in large codebases, and maintain architecture-level contextual awareness, helping developers efficiently handle multi-file projects.
- Architecture-level reasoning: The model can track the overall architecture of a codebase, understand framework dependencies, and maintain the stability of the overall architecture when making code modifications, avoiding global issues caused by local changes.
- Automatic error detection and correction: Devstral 2 can detect errors in code and attempt automatic correction, ensuring code correctness through multiple attempts and verification, and performs excellently when fixing issues in legacy systems.
- Supports multiple programming languages: The model has strong support for a variety of mainstream programming languages and can generate and optimize code in different languages according to requirements, meeting diverse development needs.
- Efficient code generation and optimization: Devstral 2 can generate high-quality, efficient code while optimizing the performance of existing code, helping developers improve development efficiency and code quality.
Devstral 2's Test Performance
- SWE-bench Verified test
- Devstral 2 (123B parameters): Accuracy reaches 72.2%, making it one of the best-performing open-source models currently available.
- Devstral Small 2 (24B parameters): Accuracy is 68.0%, with performance comparable to models five times its size.
- Comparison with competitors
- Devstral 2 vs. DeepSeek V3.2: Win rate 42.8%, clearly better than DeepSeek V3.2 (28.6%).
- When compared with Claude Sonnet 4.5 (a closed-source model): Claude Sonnet 4.5 is still more popular, indicating that there is still a gap between Devstral 2 and top closed-source models, but it is already very close.
Devstral 2 Project Address
- Project official website：https://mistral.ai/news/devstral-2-vibe-cli
- HuggingFace model repository：https://huggingface.co/collections/mistralai/devstral-2
Devstral 2 Product Pricing
-
Devstral 2 (123B parameters)：Currently available for free via API.After the free period ends, API pricing is per million input / output tokens respectively $0.40 and $2.00。
-
Devstral Small 2 (24B parameters)：API pricing is per million input / output tokens respectively $0.10 and $0.30。
Devstral 2 Application Scenarios
-
Code Generation: Devstral 2 can quickly generate code in multiple languages, helping developers start projects efficiently.
-
Code Optimization: The model supports cross-file modifications, helping developers optimize code and improve performance and maintainability.
-
Bug Fixing: The model automatically detects and fixes code errors, saving debugging time and making it suitable for legacy system repairs.
-
Enterprise Applications: The model can be customized and fine-tuned, supports architecture-level reasoning, and can meet the needs of complex enterprise projects.
-
Local Deployment: Devstral Small 2 can run locally, protecting data privacy and making it suitable for scenarios with high security requirements.