MiroThinker – An open-source series of agent models built on Qwen3

What is MiroThinker？

MiroThinker is an open-source series of intelligent agent models designed for in-depth research and solving complex, long-term problems. Built on the Qwen3 architecture, the models 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 in 8B, 14B, and 32B parameter sizes, delivering outstanding results on the GAIA benchmark. Equipped with the MiroFlow framework, the models support multilingual operations and flexible tool integration, making them suitable for a wide range of real-world applications.

Key Features of MiroThinker

Task Decomposition: Breaks down complex tasks into smaller subtasks for step-by-step resolution, improving efficiency and success rates.
Multi-hop Reasoning: Supports multi-step logical reasoning for problems that require layered thinking and information integration.
Retrieval-Augmented Generation: Combines retrieval techniques to extract relevant information from large datasets, enhancing the accuracy and richness of generated content.
Code Execution: Can directly execute code snippets, enabling solutions for programming and data processing tasks.
Web Browsing: Supports real-time web browsing to obtain up-to-date information for answering time-sensitive queries.
Document/File Processing: Reads and processes various file formats such as PDF, Word, and Excel to extract key information for analysis and answers.

Technical Principles

Qwen3-based Architecture: Built on the Qwen3 foundation model, inheriting its powerful language generation and comprehension capabilities.
Reinforcement Learning: Uses reinforcement learning to optimize behavior strategies for better performance in complex tasks. The DPO (Direct Preference Optimization) variants emphasize this optimization.
Long-term Memory and Context Management: Integrates long-term memory mechanisms to store and retrieve extensive context, supporting long-text and complex task handling.
MiroFlow Tool Integration Framework: Offers a flexible framework for seamless integration with external tools (e.g., search engines, code execution environments), extending the model’s functionality.
Large-scale Data Training: Trained on massive, high-quality datasets to ensure excellent performance across diverse task scenarios.

Project Links

GitHub Repository: https://github.com/MiroMindAI/MiroThinker
HuggingFace Model Hub: https://huggingface.co/collections/miromind-ai/mirothinker-v01-689301b6d0563321862d44a1
Online Demo: https://dr.miromind.ai/

Use Cases

Scientific Research: Assists researchers in breaking down complex scientific problems, using multi-hop reasoning and retrieval-augmented generation to propose experimental designs and solution strategies.
Business Intelligence: Gathers real-time market data, analyzes trends, and supports business decision-making to help companies maintain a competitive edge.
Education and Learning: Provides personalized learning plans and tutoring based on students’ progress and needs, improving learning outcomes.
Healthcare: Analyzes patient records and the latest medical data to offer diagnostic and treatment suggestions, aiding clinical decisions.
Intelligent Customer Support: Handles complex customer queries with multi-hop reasoning and retrieval-augmented generation to deliver accurate solutions and improve customer satisfaction.