MiroThinker - An open-source agent model series built on Qwen3

What is MiroThinker?

MiroThinker is an open-source family of intelligent agent models designed for deep research and solving complex, long-term problems. Built on Qwen3, the models feature task decomposition, multi-hop inference, retrieval-enhanced generation, code execution, web browsing, and file processing capabilities. MiroThinker v0.1 offers SFT and DPO variants with 8B, 14B, and 32B parameter scales, demonstrating excellent performance on the GAIA benchmark. The models are equipped with the MiroFlow framework, supporting multi-language models and flexible tool integration, making them suitable for a wide range of real-world applications.

MiroThinker's main functions

• Task DecompositionBreaking down complex tasks into multiple sub-tasks allows for step-by-step problem-solving, improving task execution efficiency and success rate.
• Multi-hop ReasoningIt supports multi-step logical reasoning and can handle problems that require multi-level thinking and information integration.
• Retrieval-Augmented GenerationBy combining retrieval technology, relevant information is extracted from a large amount of data, enhancing the accuracy and richness of the generated content.
• Code executionIt supports direct execution of code snippets and can handle programming tasks and data processing needs.
• Web browsingIt supports real-time web browsing to obtain the latest information, which can be used to answer questions that require the latest data.
• Document/File ProcessingIt can read and process files in various formats, such as PDF, Word, and Excel, and extract key information for analysis and answering questions.

MiroThinker's technical principles

• Architecture based on Qwen3MiroThinker is built on the Qwen3 base model, inheriting its powerful language generation and understanding capabilities.
• Reinforcement LearningThis involves optimizing the behavior of a model based on reinforcement learning to improve its performance on complex tasks. The DPO (Direct Preference Optimization) variant particularly emphasizes this point.
• Long-term memory and context managementIt integrates a long-term memory mechanism, enabling the storage and retrieval of large amounts of contextual information, and supports the processing of long texts and complex tasks.
• Tool integration framework (MiroFlow)It provides a flexible framework that supports seamless integration with external tools (such as search engines, code execution environments, etc.) to extend the functionality of the model.
• Large-scale data training: Train the model using large-scale, high-quality datasets to ensure it 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
• Experience the demo online:https://dr.miromind.ai/

Application scenarios of MiroThinker

• Scientific researchIt helps researchers break down complex scientific problems, enhances generation through multi-hop reasoning and retrieval, and provides suggestions for experimental design and solutions.
• Business IntelligenceIt provides real-time market data and trend analysis to support business decisions and help companies gain a competitive edge.
• Education and LearningWe provide personalized learning plans and tutoring based on students' learning progress and needs to improve learning outcomes.
• HealthcareIt analyzes patient medical records and the latest medical data to provide doctors with diagnostic and treatment suggestions and assist in clinical decision-making.
• Intelligent Customer ServiceIt handles complex customer inquiries by enhancing generation through multi-hop reasoning and retrieval, providing accurate solutions and improving customer satisfaction.