devstral-medium-2507

Devstral Medium 2507 is a medium-sized, code-specific language model launched by Mistral AI. Its core positioning is as a high-performance, highly generalized, enterprise-grade intelligent programming engine designed specifically for complex software development and autonomous coding agents.

• Leading SWE-Bench Verification: Achieves an accuracy rate of 61.6% in authoritative code-fixing evaluations, surpassing Gemini 2.5 Pro and GPT-4.1 while costing only one-quarter as much.
• Strong Generalization and Prompt Robustness: Delivers stable performance across diverse instruction styles, engineering environments (such as OpenHands), and toolchains, making it well-suited for a wide range of development scenarios.
• Multi-format Compatible Output: Supports both Mistral’s native function calls and XML format simultaneously, enabling flexible integration with various agent-based scaffolds.
• Full-stack Development Capability: Covers the entire development lifecycle—from code generation and test writing to dependency management, bug diagnosis, and pull request submission.

───────────────────────────────────────────────────────────────────

Core Capabilities

🧠 Deep Engineering Understanding: Can parse project structures, dependency relationships, and historical commits, precisely locating and fixing bugs in real-world repository environments.

🧩 Agent-native Architecture: As the core of a code agent, it can autonomously plan tasks, invoke CLI tools, execute tests, and verify results.

🌍 Seamless Multilingual Switching: Freely switches between dozens of languages including Python, TypeScript, and Rust, maintaining contextual consistency throughout.

🔐 Privacy and Customization: Supports on-premises deployment and domain-specific fine-tuning, enabling deep adaptation to enterprises’ internal coding standards and technology stacks.