MiroThinker
Open-Source Research Agent
MiroThinker is an open-source, search-centric research agent designed for tool-augmented reasoning and interactive scaling. Build intelligent agents that form hypotheses, retrieve evidence, and iterate in real time — just like human researchers.
A New Axis of Scaling
Beyond parameters and context — introducing Interactive Scaling
Model Parameters
Traditional approach: increase model size for better performance. Computationally expensive and diminishing returns.
Context Length
Extended context windows enable more information. Still limited by static input without iterative refinement.
Interactive Scaling
Form hypotheses, retrieve evidence via tools, revise plans based on new information, and iterate until convergence.
The MiroThinker Suite
A modular open-source ecosystem for building research agents
MiroThinker Model
Tool-native reasoning model optimized for multi-step, long-horizon tasks. Available in 30B and 235B parameter variants.
MiroFlow Framework
Orchestration framework for running, evaluating, and reproducing agent workflows with full observability.
MiroVerse Dataset
Large-scale research-agent dataset with ~147k samples designed to train search, planning, and verification behaviors.
MiroTrain Training
Infrastructure for stable agent training, including reinforcement learning and tool-use alignment pipelines.
Technical Capabilities
Built for complex, multi-step research workflows
Ingest large documents, logs, and multi-source evidence in a single reasoning session
Native support for hundreds of tool calls in complex, multi-step workflows
Achieve comparable performance to larger models with effective reasoning per FLOP
Model Versions
Choose the right variant for your use case
Research Prototype
- Initial release focused on pushing boundaries
- Higher experimental limits on tool-call depth
- Documented in accompanying arXiv paper
- Research-oriented configuration settings
Production Ready
- 30B and 235B parameter variants available
- MoE-style architecture with lower active parameters
- Major improvements in tool-call stability
- Enhanced long-horizon planning capabilities
- Cost-efficient deployment options
Built For
Real-world applications in knowledge-intensive domains
Research Automation
Automate literature review, data aggregation, and synthesis
Competitive Intelligence
Monitor markets, analyze trends, and generate insights
Technical Due Diligence
Deep-dive analysis of technical systems and codebases
Knowledge Workflows
Evidence-based Q&A and complex decision support
Open Source Community
Join developers and researchers building the future of AI agents