Devstral Small 1.1 is a software engineering-specific large language model jointly developed by Mistral AI and All Hands AI. It is fine-tuned from Mistral-Small-3.1, with its vision encoder removed to focus solely on text-based programming tasks. Below is a detailed introduction:

Technical Specifications

• Model Parameters and Context Window: Devstral Small 1.1 has 24B parameters and supports a 128k token context window, enabling it to handle extensive code files and long-context programming tasks.
• Tokenizer: It uses a custom Tekken tokenizer with a 131k vocabulary size, which helps improve the model’s understanding and processing of code-related text.
• Performance Metrics: On the SWE-bench Verified benchmark, Devstral Small 1.1 achieves a score of 53.6%, surpassing Devstral Small 1.0 by 6.8% and outperforming the second-best state-of-the-art model by 11.4%.

Key Features

• Agentic Coding Capability: As an agentic LLM, Devstral Small 1.1 excels at software engineering tasks such as codebase exploration, multi-file editing, and integration into coding agents. It can understand task requirements and generate corresponding code like an experienced developer, significantly enhancing development efficiency.
• Lightweight Design: Despite its powerful functionality, Devstral Small 1.1 is compact enough to run on consumer hardware like a single RTX 4090 or a Mac with 32GB RAM. This makes it suitable for local deployment and on-device use, offering flexibility for developers.
• Open Licensing: Released under the Apache 2.0 license, Devstral Small 1.1 allows free use and modification for both commercial and non-commercial purposes, providing developers and enterprises with greater freedom and flexibility.
• Tool Integration: The model supports Mistral’s function calling format, enabling integration with various tools to expand its capabilities and better assist developers in completing tasks.

Deployment Methods

• Via API: Create a Mistral account and obtain an API key. Follow the instructions to configure the OpenHands container and use Devstral Small 1.1 through the Mistral API.
• Local Inference: Devstral Small 1.1 can be deployed locally using libraries such as vLLM, mistral-inference, transformers, LMStudio, llama.cpp, and ollama. Among these, vLLM is recommended for implementing production-ready inference pipelines. To use it, install vLLM and mistral-common, launch the server with the command vllm serve mistralai/Devstral-Small-2507 --tokenizer_mode mistral --config_format mistral --load_format mistral --tool-call-parser mistral --enable-auto-tool-choice --tensor-parallel-size 2, and then send requests to the server using a Python script.
• Using OpenHands: OpenHands is a powerful tool for leveraging Devstral Small 1.1. First, start an OpenAI-compatible server like vLLM or Ollama. Then, use Docker to deploy OpenHands and access its UI at http://localhost:3000. In the UI, connect to the server by filling in the custom model, base URL, and API key.
• Using Cline: Similar to OpenHands, Cline can also be used with Devstral Small 1.1. Start an OpenAI-compatible server, install Cline, configure the server address, and then send requests to the server through Cline.

Use Cases

• Codebase Exploration: Devstral Small 1.1 can quickly understand and analyze codebases, helping developers familiarize themselves with new projects, locate specific code segments, and identify potential issues.
• Multi-File Editing: The model supports editing multiple files simultaneously, enabling developers to make batch modifications to code, improve code quality, and enhance development efficiency.
• Software Engineering Agent Integration: As a software engineering agent, Devstral Small 1.1 can automate various programming tasks, such as code generation, code review, and bug fixing, reducing developers’ workloads and accelerating project delivery.
• Test Coverage Analysis: When linked to a repository, Devstral Small 1.1 can analyze test coverage, identify poorly covered files, generate visualizations of test coverage, and provide explanations of the results to help developers improve testing strategies and enhance code quality.
• Game Development: Devstral Small 1.1 can assist in creating video games. For example, it can help develop a web-based game that combines elements of Space Invaders and Pong, handling tasks such as game structure design, rule definition, and code generation.

Comparison with Other Models

In the SWE-bench benchmark, Devstral Small 1.1 outperforms many other models. Under the same test framework (OpenHands), it surpasses larger models like Deepseek-V3-0324 and Qwen3 232B-A22B. This demonstrates its strong capabilities in software engineering tasks.

Devstral Small 1.1 is a powerful tool for software engineering tasks. Its lightweight design, strong agentic coding capabilities, and strong performance make it an excellent assistant for developers. Whether used via API or local deployment, it can adapt to different development scenarios and help developers improve efficiency and code quality. For enterprises requiring specialized capabilities, Mistral AI also plans to release commercial models to meet specific needs.