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:
- π€ 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.
- π€ MiroThinker: an open-source agent that natively supports tool-assisted reasoning. See MiroThinker.
- π MiroVerse: 147k premium open-source training data supporting research agent training. See MiroVerse.
- π° News & Updates
- π Get Started in Under 5 Minutes
- π€ What is MiroFlow?
- π Highlights
- π Performance on Benchmarks
- π§ Supported Models & Tools
- β FAQ
- π€ Contributing
- π License
- π Acknowledgments
- [2025-09-15]: ππ 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.
- Python: 3.12 or higher
- Package Manager:
uv - Operating System: Linux, macOS
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"π Expected Output: Your agent should return \boxed{Congo Democratic Republic} π
π‘ Tip: If you encounter issues, check that your API key is correctly set in the
.envfile and that all dependencies are installed.
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 AdviceDemo-Research-Assistant.mp4 |
- 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.
- Models: GPT, Claude, Gemini, Qwen, etc.
- Tools: Audio Transcription, Python, File Reading, Reasoning, Google Search, VQA, E2B, etc.
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
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.
We welcome contributions from the community! Whether you're fixing bugs, adding features, or improving documentation, your help is appreciated.
- π Issues: Report bugs or request features via GitHub Issues.
- π Pull Requests: Submit improvements via pull requests.
- π¬ Discussions: Join our Discord community for questions and discussions.
This project is licensed under the Apache License 2.0.
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!
@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}
}