MiroMind Research Agent v0.1
MiroMind team
Aug 7, 2025
MiroMind Research Agent v0.1: A High-Performance, Fully Open-Sourced Research Agent Project That Grows with Developers
Deep Research Demo is live - Try Now
What is Research Agent?
A research agent is a system that integrates LLMs/VLMs with an agent framework, capable of conducting multi-step research on the internet to tackle complex tasks. It can accomplish in just tens of minutes what would otherwise take humans many hours to complete.
The Goal of MiroMind Research Agent Project
MiroMind Research Agent Project is to build a fully open-sourced, high-performance research agent ecosystem — with an agent framework, model, data, and training infra all fully accessible and open. We invite developers to participate together to promote the healthy growth of the research agent ecosystem.
MiroMind Research Agent vs. Existing Research Agents (2025-08-08)
Fully Open-Sourced
Compared to existing methods, the MiroMind Research Agent opens up the entire stack of research agents, including four sub-projects: MiroFlow (agent framework), MiroThinker (model), MiroVerse (data), and MiroTrain / MiroRL (training infra).
Project | GAIA validation Performance | Open Source Scope | Reproducible with open-source suites | ||||
|---|---|---|---|---|---|---|---|
Technical Report / Blog | Agent Framework | Model | Data | Training Framework | |||
MiroMind Research Agent | 82.4 | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ |
Open AI Deep Research 20250202 | 67.4 | ✔️ | - | - | - | - | - |
Manus | 73.3 | ✔️ | - | - | - | - | - |
Aworld | 61.8 | ✔️ | - | - | - | - | - |
OWL | 69.1 | ✔️ | ✔️ | ✔️ | - | - | - |
WebShaper | 60.9 | ✔️ | ✔️ | ✔️ | - | - | - |
Details of MiroMind Research Agent
Sub-Project | Type | Description | Highlights of This Version | Open Source Link |
|---|---|---|---|---|
MiroFlow v0.1 | Agent Framework | Supporting multiple mainstream tool calls, extending large language models to support tool-assisted research agent reasoning | Stable reproduction, breakthrough performance of 82.4 on GAIA, currently the highest reproducible performance model on GAIA | |
MiroThinker v0.1 | Model | Large language model that natively supports tool-assisted reasoning | Trainable and reproducible, top performance in GAIA | https://github.com/MiroMindAI/MiroThinker https://huggingface.co/collections/miromind-ai/mirothinker-v01-689301b6d0563321862d44a1 |
MiroVerse v0.1 | Data | 147k Open-Source training data supporting research agent training | Focus on community feedback, monthly update continuously provide monthly updates with high-quality research agent datasets. | |
MiroTrain v0.1 / MiroRL v.01 | Training Infra | Supports stable and efficient training for the research agent models | Covers the whole research agent training process, supports long-text training and RL training with tools | https://github.com/MiroMindAI/MiroRL |
Reproducible State-of-the-Art Performance
MiroFlow achieves state-of-the-art performance among open-source projects on four comprehensive leaderboards — GAIA Val (82.4%), HLE (27.2%), xBench (72.0%), and BrowseComp-ZH (47.1%) — with performance approaching that of top closed-source agents! Every reported metric is fully reproducible with the publicly available code: github.com/MiroMindAI/MiroFlow.
MiroThinker is an open-source research agent model that achieved state-of-the-art open-source performance (60.2%) on the GAIA-Text-103 benchmark, narrowing the gap with OpenAI Deep Research.
In addition, by combining MiroFlow and MiroThinker, research agent services can run on a single RTX 4090 GPU. The entire technology stack relies on free and open-source tools, making deployment, scaling, and reproduction simple. Code repository: https://github.com/MiroMindAI/MiroThinker
Method | Open-Source Free Tool Set | GAIA Text-103 Best Pass@1 | GAIA Text-103 Pass@1 (Avg@8) | GAIA Val Best Pass@1 | GAIA Val Pass@1 (Avg@8) |
|---|---|---|---|---|---|
- | - | - | 67.4 | - | |
MiroThinker-8B | ✔️ | 46.6 | 44.8 | 37.0 | 35.4 |
MiroThinker-8B | - | 50.5 | 46.7 | 38.2 | 35.9 |
MiroThinker-32B | ✔️ | 57.3 | 54.1 | 48.5 | 45.9 |
MiroThinker-32B | - | 60.2 | 57.9 | 50.9 | 48.9 |
What Do We Offer For the Developer Community?
A fully open-sourced, reproducible research agent project, easy to learn, reuse, and contribute to.
A co-constructive community environment with monthly releases, where every developer’s suggestions and feedback are prioritized to directly drive project updates and feature evolution.
Broad participation and customization opportunities — whether you are a data engineer, AI researcher, or a developer enthusiastic about new tools, you can initiate challenges, contribute data, or customize Agents.
Rich community activities: Regular fine-tuning competitions, SOTA leaderboards, behavior markets, hackathons, and diverse gameplay to make participation rewarding.
About Us
MiroMind is creating the self-aware digital being, evolving together with the community to unlock safe, beneficial AGI for all. We believe: "AGI is never a perfect model but a constantly evolving system." At Miromind, you are not just a “user” but a co-builder. Every step of our growth and every update is inseparable from your feedback. Your growth is the core of Miromind’s progress.