MiroThinker is an open-source, research-grade search agent from MiroMindAI optimized for tool-augmented reasoning and deep information seeking. It ships as a complete ecosystem including the MiroThinker model, the MiroFlow agent framework, the MiroVerse training dataset, and supporting infrastructure.
Extended Context and Tool Calling
- Very long context windows of up to 256K tokens maintain coherence across extended document traces and prolonged multi-step reasoning chains
- High-frequency tool calling support handles hundreds to thousands of sequential invocations with robust trace collection and logging
- Full reproducibility ensured through centralized citation management and comprehensive evaluation logging
Complete Research Ecosystem
- Bundles models, a reproducible agent framework (MiroFlow), curated datasets (MiroVerse), and benchmark suites
- Researchers can evaluate and compare performance systematically across controlled experiment configurations
- Configurable tool integrations include web search, code execution, summarization, and scrapers
Retrieval and Deployment
- Retrieval pipelines combine hybrid search with re-ranking for high-accuracy information retrieval
- Docker-friendly deployment and multiple serving options for local or cloud environments
- Implemented primarily in Python with clear, modular component separation for easy extension