What is MiroThinker
MiroThinker is an open-source series of agent models designed for deep research and complex, long-horizon problem solving. The models are built on Qwen3 and feature capabilities such as task decomposition, multi-hop reasoning, retrieval-augmented generation, code execution, web browsing, and file processing. MiroThinker v0.1 offers SFT and DPO variants at 8B, 14B, and 32B parameter scales, and performs well on the GAIA benchmark. The models come with the MiroFlow framework, support multilingual models and flexible tool integration, and are suitable for a wide range of real-world application scenarios.
MiroThinker's Main Features
- Task Decomposition: Breaks complex tasks into multiple subtasks, solves problems step by step, and improves the efficiency and success rate of task execution.
- Multi-hop Reasoning: Supports multi-step logical reasoning and can handle problems that require multiple layers of thinking and information integration.
- Retrieval-Augmented Generation: Combines retrieval techniques to extract relevant information from large amounts of data, enhancing the accuracy and richness of generated content.
- Code Execution: Supports direct execution of code snippets and can handle needs such as programming tasks and data processing.
- Web Browsing: Supports real-time web browsing to obtain the latest information, and is used to answer questions that require up-to-date data.
- Document/File Processing: Can read and process files in multiple formats, such as PDF, Word, Excel, etc., and extract key information for analysis and answering questions.
MiroThinker's Technical Principles
- Architecture Based on Qwen3: MiroThinker is built on the Qwen3 foundation model and inherits its powerful language generation and understanding capabilities.
- Reinforcement Learning: Optimizes the model's behavior policy based on reinforcement learning, enabling better performance in complex tasks. The DPO (Direct Preference Optimization) variant particularly emphasizes this.
- Long-term Memory and Context Management: Integrates a long-term memory mechanism that can store and retrieve large amounts of contextual information and support the processing of long texts and complex tasks.
- Tool Integration Framework (MiroFlow): Provides a flexible framework that supports seamless integration with external tools (such as search engines, code execution environments, etc.) to extend the model's capabilities.
- Large-scale Data Training: Trained on large-scale, high-quality datasets to ensure the model performs well across a variety of task scenarios.
MiroThinker's Project Address
- GitHub Repository：https://github.com/MiroMindAI/MiroThinker
- HuggingFace Model Library：https://huggingface.co/collections/miromind-ai/mirothinker-v01-689301b6d0563321862d44a1
- Online Experience Demo：https://dr.miromind.ai/
MiroThinker's Application Scenarios
- Scientific Research: Helps researchers decompose complex scientific problems and, through multi-hop reasoning and retrieval-augmented generation, provides suggestions for experimental design and solutions.
- Business Intelligence: Obtains market data in real time, analyzes trends, and provides support for business decisions, helping enterprises gain an advantage in competition.
- Education and Learning: Provides personalized learning plans and tutoring based on students' learning progress and needs, improving learning outcomes.
- Healthcare: Analyze patient medical records and the latest medical data to provide doctors with diagnostic and treatment recommendations, assisting clinical decision-making.
- Intelligent Customer Service: Handle complex customer inquiries, and through multi-hop reasoning and retrieval-augmented generation, provide accurate solutions and improve customer satisfaction.