Tongyi DeepResearch is an open-source web agent that matches the performance of OpenAI’s DeepResearch on several benchmarks. It scored 32.9 on Humanity’s Last Exam (HLE), 43.4 on BrowseComp, and 46.7 on BrowseComp-ZH, excelling in complex information-seeking tasks. The model also achieved a score of 75 on the xbench-DeepSearch benchmark, outperforming existing deep research agents. The creators provide a comprehensive methodology for developing such advanced agents, including a novel data synthesis solution that covers the entire training process. The training involves Agentic Continual Pre-training (CPT) and Supervised Fine-Tuning (SFT), culminating in a Reinforcement Learning (RL) phase. The data collection process is extensive, pulling from various sources like documents and knowledge graphs. Multi-style question-answer pairs are generated from this data, with first-order and higher-order action synthesis data enhancing decision-making capabilities. A fully automated process produces high-quality synthetic datasets without human intervention, allowing for the creation of complex, multi-source questions and answers through a structured pipeline. The model incorporates two key reasoning frameworks: the ReAct framework for multi-turn reasoning and the IterResearch paradigm, which creates a streamlined workspace for each decision. This enhances the agent's ability to plan and use tools effectively, particularly for long-horizon tasks. The agent operates in two modes: the native ReAct Mode, which follows a straightforward Thought-Action-Observation cycle, and the Heavy Mode, designed for more intricate, multi-step research tasks. Both modes leverage scalable computation methods to maximize the model's potential, moving beyond traditional, complex designs.