Mistral AI Introduces Devstral2507 Series for Developers

Mistral AI, in partnership with All Hands AI, has launched the Devstral2507 series, featuring two advanced large language models (LLMs) designed for code reasoning, program synthesis, and agent-based task execution. The release includes Devstral Small1.1 (open-source) and Devstral Medium2507 (enterprise-only), both optimized for performance and cost-efficiency in development tools and automation systems.

Open-Source Power: Devstral Small1.1

Devstral Small1.1 is built on the Mistral-Small-3.1 base model, boasting 24 billion parameters and a 128k context window. This enables seamless processing of multi-file code inputs and complex prompts, aligning with real-world software engineering workflows. The model is fine-tuned for structured outputs (e.g., XML, function calls) and integrates with frameworks like OpenHands for tasks such as code navigation and multi-step editing. Licensed under Apache 2.0, it supports both research and commercial use.

In benchmark testing, Devstral Small1.1 achieved 53.6% accuracy on the SWE-Bench Verified test, demonstrating robust performance in generating patches for GitHub issues. While it trails behind large commercial models, its balance of size, cost, and inference capability makes it ideal for local development and client-side tools.

Enterprise-Grade Performance: Devstral Medium2507

The closed-source Devstral Medium2507, available via Mistral’s API or enterprise agreements, outperforms its smaller counterpart with a 61.6% SWE-Bench score. It excels in long-context reasoning, surpassing models like Gemini 2.5 Pro and GPT-4.1 in code repository tasks. Though pricier, its precision in structured code editing makes it a top choice for high-performance production environments.

Deployment Flexibility

• Devstral Small1.1: Offered in quantized formats for local use on RTX4090 GPUs or Apple Silicon machines (32GB+ RAM). Also accessible via Mistral’s API at rates matching the Mistral-Small series.
• Devstral Medium2507: Exclusively cloud-hosted or enterprise-deployed, catering to large-scale code automation needs.

Both models support agent framework integration, streamlining workflows for test generation, refactoring, and error resolution.

Key Points

• 🌟 Dual Offerings: Open-source Small1.1 and enterprise Medium2507 cater to diverse coding needs.
• 🚀 Benchmark Excellence: Small1.1 (53.6%) and Medium2507 (61.6%) outperform peers in code tasks.
• 💼 Scalable Use Cases: From local experimentation to enterprise-grade automation.