MiroMindAI/MiroFlow

---

This repo is the official implementation of the MiroMind Research Agent Project. It is a leading-performance, fully open-source system designed to perform multi-step internet research for addressing complex challenges such as future event prediction. The project currently comprises four key components:

• 🤖 MiroFlow: an open-source research agent framework that offers reproducible state-of-the-art performance on representative benchmarks (e.g., FutureX, GAIA, HLE, xBench-DeepSearch, and BrowserComp benchmarks), included in this repo. See [Get Started in Under 5 Minutes] for a quick start.
• 🤔 MiroThinker: an open-source agent that natively supports tool-assisted reasoning. See MiroThinker.
• 📊 MiroVerse: 147k premium open-source training data supporting research agent training. See MiroVerse.

---

📋 Table of Contents

• 📰 News & Updates
• 🚀 Get Started in Under 5 Minutes
• 🤖 What is MiroFlow?
• 🌟 Highlights
• 📈 Performance on Benchmarks
• 🔧 Supported Models & Tools
• ❓ FAQ
• 🤝 Contributing
• 📄 License
• 🙏 Acknowledgments

---

📰 News & Updates

• [2025-09-15]: 🎉🎉 MiroFlow v0.3: Enhanced codebase architecture and significantly improved benchmark performance, boosting GPT-5's prediction accuracy for future events by 11%. MiroFlow now ranks #1 in the future prediction benchmark. See FutureX.
• [2025-08-27]: MiroFlow v0.2: Achieves state-of-the-art performance across multiple agentic benchmarks, including HLE (27.2%), HLE-Text-Only (29.5%), BrowserComp-EN (33.2%), BrowserComp-ZH (47.1%), and xBench-DeepSearch (72.0%).
• [2025-08-26]: Released GAIA Validation Trace (73.94% pass@1) and Gradio Demo for local deployment.
• [2025-08-08]: MiroFlow v0.1: Complete open-source release of the research agent framework.

---

🚀 Get Started in Under 5 Minutes

📋 Prerequisites

• Python: 3.12 or higher
• Package Manager: uv
• Operating System: Linux, macOS

⚡ Quick Setup

Example: Intelligent document analysis with file processing capabilities.

# 1. Clone and setup
git clone https://github.com/MiroMindAI/MiroFlow && cd MiroFlow
uv sync

# 2. Configure API key
cp .env.template .env
# Edit .env and add your OPENROUTER_API_KEY

# 3. Run your first agent
uv run main.py trace --config_file_name=agent_quickstart_reading --task="What is the first country listed in the XLSX file that have names starting with Co?" --task_file_name="data/FSI-2023-DOWNLOAD.xlsx"

🎉 Expected Output: Your agent should return \boxed{Congo Democratic Republic} 😊

💡 Tip: If you encounter issues, check that your API key is correctly set in the .env file and that all dependencies are installed.

---

🤖 What is MiroFlow?

MiroFlow is a high-performance, modular framework for building intelligent AI agents that deliver state-of-the-art results on complex reasoning tasks like future event prediction. The framework features advanced multi-turn conversation capabilities, extensive tool ecosystem integration, and hierarchical sub-agent orchestration for optimal task completion. Learn more about our agent framework.

Research Assistant Demo - Read CVPR 2025 Best Paper and Provide Research Advice

---

🌟 Highlights

• Reproducible State-of-the-Art Performance: #1 ranking across multiple representative agentic benchmarks, including FutureX, GAIA, HLE, xBench-DeepSearch, and BrowserComp benchmarks)
• High Concurrency & Reliability: Built with robust concurrency management and fault-tolerant design, MiroFlow efficiently handles rate-limited APIs and unstable networks, ensuring seamless trajectory collection and reliable execution of complex tasks.
• Cost-Effective Deployment: Powered by the open-source MiroThinker model, MiroFlow can run a research agent service on a single RTX 4090. The entire stack relies on free, open-source tools, making it simple to deploy, scale, and reproduce. See MiroThinker.

---

🔧 Supported Models & Tools

• Models: GPT, Claude, Gemini, Qwen, etc.
• Tools: Audio Transcription, Python, File Reading, Reasoning, Google Search, VQA, E2B, etc.

---

📈 Performance on Benchmarks

We achieved the #1 ranking on the FutureX Benchmark Leaderboard as of September 10, 2025, boosting GPT-5's prediction accuracy for future events by 11%.

We benchmark MiroFlow on a series of benchmarks, including GAIA, HLE, BrowseComp, and xBench-DeepSearch, and achieved SOTA results.

Model/Framework	GAIA Val	HLE	HLE-Text	BrowserComp-EN	BrowserComp-ZH	xBench-DeepSearch
MiroFlow	82.4%	27.2%	29.5%	33.2%	47.1%	72.0%
OpenAI Deep Research	67.4%	26.6%	-	51.5%	42.9%	-
Gemini Deep Research	-	26.9%	-	-	-	50+%
Kimi Researcher	-	-	26.9%	-	-	69.0%
WebSailor-72B	55.4%	-	-	-	30.1%	55.0%
Manus	73.3%	-	-	-	-	-
DeepSeek v3.1	-	-	29.8%	-	-	71.2%

Follow our detailed guides to reproduce benchmark results in our Benchmarks Documentation

---

❓ FAQ

What API keys do I need?

You only need an OpenRouter API key to get started. OpenRouter provides access to multiple language models through a single API.

Can I use other language models besides OpenRouter?

Yes, MiroFlow supports various language models. Check our documentation for configuration details.

How do I reproduce the benchmark results?

Follow our detailed Benchmarks Documentation for step-by-step reproduction guides.

Is there commercial support available?

For commercial inquiries and enterprise support, please contact us through our website.

---

🤝 Contributing

We welcome contributions from the community! Whether you're fixing bugs, adding features, or improving documentation, your help is appreciated.

• 📋 Issues: Report bugs or request features via GitHub Issues.
• 🔀 Pull Requests: Submit improvements via pull requests.
• 💬 Discussions: Join our Discord community for questions and discussions.

📄 License

This project is licensed under the Apache License 2.0.

🙏 Acknowledgments

Benchmark Contributors for the comprehensive evaluation datasets.

Open Source Community for the tools and libraries that make this possible.

We thank all contributors who have helped make MiroFlow better:

Join our community and help us build the future of AI agents!

References

@article{miromind2025mirothinker,
  title={MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling},
  author={MiroMind Team and Bai, Song and Bing, Lidong and Chen, Carson and Chen, Guanzheng and Chen, Yuntao and Chen, Zhe and Chen, Ziyi and Dong, Xuan and others},
  journal={arXiv preprint arXiv:2511.11793},
  year={2025}
}