Tongyi DeepResearch Technical Report

  • 2025-10-28 17:53:02
  • Tongyi DeepResearch Team, Baixuan Li, Bo Zhang, Dingchu Zhang, Fei Huang, Guangyu Li, Guoxin Chen, Huifeng Yin, Jialong Wu, Jingren Zhou, Kuan Li, Liangcai Su, Litu Ou, Liwen Zhang, Pengjun Xie, Rui Ye, Wenbiao Yin, Xinmiao Yu, Xinyu Wang, Xixi Wu, Xuanzhong Chen, Yida Zhao, Zhen Zhang, Zhengwei Tao, Zhongwang Zhang, Zile Qiao, Chenxi Wang, Donglei Yu, Gang Fu, Haiyang Shen, Jiayin Yang, Jun Lin, Junkai Zhang, Kui Zeng, Li Yang, Hailong Yin, Maojia Song, Ming Yan, Peng Xia, Qian Xiao, Rui Min, Ruixue Ding, Runnan Fang, Shaowei Chen, Shen Huang, Shihang Wang, Shihao Cai, Weizhou Shen, Xiaobin Wang, Xin Guan, Xinyu Geng, Yingcheng Shi, Yuning Wu, Zhuo Chen, Zijian Li, Yong Jiang
  • 0

Abstract

We present Tongyi DeepResearch, an agentic large language model, which isspecifically designed for long-horizon, deep information-seeking researchtasks. To incentivize autonomous deep research agency, Tongyi DeepResearch isdeveloped through an end-to-end training framework that combines agenticmid-training and agentic post-training, enabling scalable reasoning andinformation seeking across complex tasks. We design a highly scalable datasynthesis pipeline that is fully automatic, without relying on costly humanannotation, and empowers all training stages. By constructing customizedenvironments for each stage, our system enables stable and consistentinteractions throughout. Tongyi DeepResearch, featuring 30.5 billion totalparameters, with only 3.3 billion activated per token, achievesstate-of-the-art performance across a range of agentic deep researchbenchmarks, including Humanity's Last Exam, BrowseComp, BrowseComp-ZH,WebWalkerQA, xbench-DeepSearch, FRAMES and xbench-DeepSearch-2510. Weopen-source the model, framework, and complete solutions to empower thecommunity.

 

Quick Read (beta)

loading the full paper ...