An open-source deep research agent optimized for research and prediction tasks, achieving 80.8% Avg@8 score on the challenging GAIA benchmark, featuring 256K context window support and up to 600 tool calls per task.
One-Minute Overview#
MiroThinker is an open-source deep research agent specifically designed for research and prediction tasks. It demonstrates exceptional performance across multiple authoritative benchmarks including GAIA and BrowseComp, achieving state-of-the-art results. MiroThinker supports an impressive 256K context window and high-frequency tool usage, enabling it to handle complex long-term reasoning and multi-step analysis tasks.
Core Value: Through "interactive scaling" technology, it introduces a third dimension of performance improvement beyond model size and context length, significantly enhancing the deep analysis capabilities of research agents.
Quick Start#
Installation Difficulty: Medium - The project provides comprehensive deployment guides and requires Python environment with relevant dependencies, but offers multiple deployment options to suit different needs.
# Clone the repository
git clone https://github.com/MiroMindAI/MiroThinker.git
# Install dependencies
pip install -r requirements.txt
Is this suitable for my scenario?
- ✅ Deep Research Requirements: Ideal for research tasks requiring long-term, multi-step analysis such as market research, academic research
- ✅ Predictive Analytics: Excels in areas like financial prediction, surpassing Kimi-K2-Thinking on BrowseComp-ZH
- ❌ Simple Q&A: May be overly complex and resource-intensive for single-turn simple questions
- ❌ Resource-Constrained Environments: Large parameter versions require significant computational resources, not suitable for small-scale personal deployments
Core Capabilities (Optional)#
1. Interactive Scaling - Overcoming Traditional Research Agent Limitations#
- Trains agents to handle deeper and more frequent agent-environment interactions as a third dimension of performance improvement, going beyond just scaling model size or context length Actual Value: Significantly enhances the analysis depth and accuracy of agents, enabling them to handle more complex research tasks
2. 256K Context Window - Supporting Long-Term Reasoning#
- Supports ultra-long context windows capable of handling complex research tasks requiring extensive background information Actual Value: Maintains analysis continuity without frequent context truncation, suitable for deep report generation
3. High-Frequency Tool Usage - Enhanced Analysis Capabilities#
- Supports up to 600 tool calls per task (v1.0) or 400 (v1.5), far exceeding previous open-source research agents Actual Value: Can acquire and integrate large amounts of multi-source information, improving the comprehensiveness and accuracy of research analysis
4. Multi-Scale Model Options - Flexible Adaptation to Different Needs#
- Offers models ranging from 8B to 235B parameters, adapting to different computational budgets and application scenarios Actual Value: Users can choose appropriate model versions based on their resources and needs, balancing performance with cost
Tech Stack & Integration#
Development Language: Python Key Dependencies: Based on Qwen series models, including various training and optimization tools Integration Method: SDK/Library - Provides complete tools and framework support, seamless integration with external tools and APIs
Maintenance Status (Optional)#
- Development Activity: Highly active - Continuously iterating since initial release in August 2025, now at v1.5
- Recent Updates: Recent - Added research report generation and multiple document upload support in January 2026
- Community Response: Active - Project continues to receive performance improvements and new feature updates
Commercial & Licensing (Optional)#
License: Open source license
- ✅ Commercial: Commercial use allowed
- ✅ Modification: Modification and distribution allowed
- ⚠️ Restrictions: Specific restrictions subject to project license file
Documentation & Learning Resources (Optional)#
- Documentation Quality: Comprehensive - Includes detailed technical reports, quick start guides, and FAQs
- Official Documentation: Project README contains complete usage guides and performance evaluations
- Sample Code: Provides Gradio demo and multiple deployment options