MiroThinker is an open-source series of agentic large language models designed for deep research and complex, long-horizon problem-solving. Built upon the Qwen3 architecture, it offers models in 8B, 14B, and 32B parameter sizes, with both Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO) variants. MiroThinker excels in tasks requiring tool use, code execution, web browsing, and document processing, demonstrating state-of-the-art performance among open-source models on benchmarks like GAIA.

How It Works

MiroThinker integrates advanced capabilities such as task decomposition, multi-hop reasoning, and retrieval-augmented generation. It leverages the MiroFlow framework, which provides a robust environment for agent development, featuring enhanced conversation management, flexible tool integration (supporting both open-source and commercial tools), and comprehensive benchmark evaluations. The models are trained on the MiroVerse dataset and utilize the MiroTrain framework for efficient training.

Quick Start & Requirements

• Installation: Clone the repository, download benchmark data (password: pf4*), and set up the environment using uv sync.
• Prerequisites: Python 3.10+, uv package manager, and API keys for various services (Serper, E2B, OpenAI, Anthropic, etc., depending on tool configuration).
• Setup: Requires configuring API keys in a .env file. Serving models can be done via SGLang or quantized methods (llama.cpp, Ollama).
• Links:
  • Demo: https://dr.miromind.ai/
  • Models: https://huggingface.co/collections/miromind-ai/mirothinker-v01-689301b6d0563321862d44a1
  • Data: https://huggingface.co/datasets/miromind-ai/MiroVerse-v0.1
  • Deployment Docs: USE-OS-TOOL.md (referenced in README)

Highlighted Details

• Achieves state-of-the-art performance on the GAIA benchmark among open-source models.
• Offers SFT and DPO variants across 8B, 14B, and 32B parameter scales.
• Supports both open-source and commercial tools for enhanced capabilities.
• Includes a comprehensive benchmark suite covering GAIA, HLE, BrowseComp, and more.

Maintenance & Community

• Community: Discord server available (https://discord.com/invite/GPqEnkzQZd).
• Updates: Recent updates include light-weight deployment options and the release of v0.1 models, framework, and data.

Licensing & Compatibility

• License: Apache License 2.0.
• Compatibility: Permissive license suitable for commercial use and integration with closed-source projects.

Limitations & Caveats

The model's Chinese language capabilities are currently limited due to the predominantly English nature of the MiroVerse-v0.1 dataset, with plans to improve this in future versions. Performance metrics are reported using both "Best Pass@1" and "Pass@1 (Avg@8)" for stability and peak performance comparison.