MiroThinker
NormalDescription
A deep research agent framework optimized for complex research and prediction tasks, with MiroThinker-1.7 and MiroThinker-H1 models achieving 74.0 and 88.2 on BrowseComp benchmark, supporting multi-step reasoning and information retrieval.
Key Features
- Deep research agent framework optimized for complex research and prediction tasks
- MiroThinker-1.7 achieves 74.0 on BrowseComp, H1 model reaches 88.2
- Supports 256K context window and up to 300 tool calls per task
- Available in 30B and 235B parameter scales for different compute budgets
- Interactive Scaling as a third dimension of performance improvement
- Open-source SOTA on BrowseComp-ZH at 1/30 the cost of proprietary models
Use Cases
💡 Deep web research and multi-step information retrieval tasks
💡 Financial prediction and market trend analysis
💡 Complex Q&A requiring large-context reasoning
💡 Automated literature review and research report generation
💡 Building high-performance search-augmented AI applications
Strengths & Limitations
✅ Strengths
- • High community interest (8.4k stars)
- • Permissive open-source license (Apache-2.0)
- • Responsive to issues, low backlog
Categories
Quick Start
Visit dr.miromind.ai for online demo. Clone the repository and install dependencies per README. Download MiroThinker-1.7 model weights (30B or 235B). Launch the research agent using the provided tools and workflows. Test interactively via the Gradio demo.