Detailed Introduction
MiroThinker, from MiroMindAI, is an open-source research-grade search agent and framework focused on tool-augmented reasoning and deep information seeking. The project includes the model (MiroThinker), an agent framework (MiroFlow), a dataset (MiroVerse), and training infrastructure. It supports very long contexts (up to 256K) and hundreds-to-thousands of tool calls, enabling complex research workflows. See the project homepage at miromind.ai and try the interactive demo at dr.miromind.ai.
Main Features
- Open research search agent designed for tool usage and multi-step reasoning.
- Very long context windows (up to 256K) for handling long documents and extended traces.
- High-frequency tool calling support with robust trace collection and logging.
- Full ecosystem: models, reproducible agent framework, datasets, and benchmark suites for evaluation.
Use Cases
Suited for academic research, long-document Q&A, deep web retrieval, benchmark evaluation, and developer experimentation. Researchers can reproduce benchmark results and run evaluations; engineering teams can integrate MiroThinker as a tool-augmented retrieval or research assistant subsystem.
Technical Features
Implemented primarily in Python, MiroThinker provides a configurable agent framework with tool integrations (web search, code execution, summarization, scrapers), Docker-friendly deployment, and multiple serving options. Retrieval pipelines include hybrid search, re-ranking, and centralized citation management to preserve reproducibility and traceability in evaluations.