A couple of agentic LLMs for software engineering tasks, excelling at using tools to explore codebases, edit multiple files, and power SWE Agents. • 2 items • Updated • 56
Devstral 2 123B Instruct 2512
Devstral is an agentic LLM for software engineering tasks. Devstral 2 excels at using tools to explore codebases, editing multiple files and power software engineering agents.
The model achieves remarkable performance on SWE-bench.
This model is an Instruct model in FP8, fine-tuned to follow instructions, making it ideal for chat, agentic and instruction based tasks for SWE use cases.
For enterprises requiring specialized capabilities (increased context, domain-specific knowledge, etc.), we invite companies to reach out to us.
Key Features
The Devstral 2 Instruct model offers the following capabilities:
- Agentic Coding: Devstral is designed to excel at agentic coding tasks, making it a great choice for software engineering agents.
- Improved Performance: Devstral 2 is a step-up compared to its predecessors.
- Better Generalization: Generalises better to diverse prompts and coding environments.
- Context Window: A 256k context window.
Use Cases
AI Code Assistants, Agentic Coding, and Software Engineering Tasks. Leveraging advanced AI capabilities for complex tool integration and deep codebase understanding in coding environments.
Benchmark Results
| Model/Benchmark | Size (B Parameters) | SWE Bench Verified | SWE Bench Multilingual | Terminal Bench 2 |
|---|---|---|---|---|
| Devstral 2 | 123 | 72.2% | 61.3% | 32.6% |
| Devstral Small 2 | 24 | 68.0% | 55.7% | 22.5% |
| GLM 4.6 | 355 | 68.0% | -- | 24.6% |
| Qwen 3 Coder Plus | 480 | 69.6% | 54.7% | 25.4% |
| MiniMax M2 | 230 | 69.4% | 56.5% | 30.0% |
| Kimi K2 Thinking | 1000 | 71.3% | 61.1% | 35.7% |
| DeepSeek v3.2 | 671 | 73.1% | 70.2% | 46.4% |
| GPT 5.1 Codex High | -- | 73.7% | -- | 52.8% |
| GPT 5.1 Codex Max | -- | 77.9% | -- | 60.4% |
| Gemini 3 Pro | -- | 76.2% | -- | 54.2% |
| Claude Sonnet 4.5 | -- | 77.2% | 68.0% | 42.8% |
*Benchmark results presented are based on publicly reported values for competitor models.
Usage
Scaffolding
Together with Devstral 2, we are releasing Mistral Vibe, a CLI tool allowing developers to leverage Devstral capabilities directly in your terminal.
- Mistral Vibe (recommended): Learn how to use it here
Devstral 2 can also be used with the following scaffoldings:
You can use Devstral 2 either through our API or by running locally.
Mistral Vibe
The Mistral Vibe CLI is a command-line tool designed to help developers leverage Devstral’s capabilities directly from their terminal.
We recommend installing Mistral Vibe using uv for faster and more reliable dependency management:
uv tool install mistral-vibe
You can also run:
curl -LsSf https://mistral.ai/vibe/install.sh | sh
If you prefer using pip, use:
pip install mistral-vibe
To launch the CLI, navigate to your project's root directory and simply execute:
vibe
If this is your first time running Vibe, it will:
- Create a default configuration file at
~/.vibe/config.toml. - Prompt you to enter your API key if it's not already configured, follow these instructions to create an Account and get an API key.
- Save your API key to
~/.vibe/.envfor future use.
Local Deployment
The model can also be deployed with the following libraries, we advise everyone to use the Mistral AI API if the model is subpar with local serving:
vllm (recommended): See heresglang: See heretransformers: See here
We're thankful to the llama.cpp team and their community as well as the LM Studio and Ollama teams that worked hard to make these models also available for their frameworks.
You can now also run Devstral using these (alphabetical ordered) frameworks:
llama.cpp: To use community ones such as Unsloth's or Bartowski's make sure to use changes from this PR.Ollama: https://ollama.com/library/devstral-2
If you notice subpar performance with local serving, please submit issues to the relevant framework so that it can be fixed and in the meantime we advise to use the Mistral AI API.
vLLM (recommended)
SGLang
Transformers
Tests
To help test our model via vLLM or test that other frameworks' implementations are correct, here is a set of prompts you can try with the expected outputs.
- Call one tool
- Call tools one at a time subsequently
- Long context
- Chatting tech
- Small talk
Run the examples above with the following python script which assumes there is an OpenAI compatible server deployed at localhost:8000:
License
This model is licensed under a Modified MIT License.
You must not use this model in a manner that infringes, misappropriates, or otherwise violates any third party’s rights, including intellectual property rights.
- Downloads last month
- 7,644
Safetensors
Model size
125B params
Tensor type
BF16
·
F8_E4M3 ·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 23 Ask for provider support