On the morning of September 17th, Alibaba open-sourced its first deep research Agent model - Tongyi DeepResearch. The model achieved state-of-the-art (SOTA) performance across multiple authoritative benchmark datasets including HLE, BrowseComp-zh, and GAIA, surpassing Agent models such as OpenAI Deep Research and DeepSeek-V3.1. Currently, Tongyi DeepResearch's model, framework, and solutions are fully open-sourced, with users able to download the model and code from Github, Hugging Face, and ModelScope community.

It is reported that existing deep research paradigms face challenges of "cognitive space suffocation" and "irreversible noise pollution" when processing long-cycle tasks, leading to degraded reasoning capabilities and ultimately making it difficult to complete truly long-term, complex research tasks. The Tongyi team constructed a complete training pipeline driven by synthetic data that spans pre-training and post-training phases, significantly improving the model's iteration speed and generalization capabilities.

On authoritative Agent benchmark datasets including Humanity's Last Exam (HLE), BrowseComp, BrowseComp-ZH, GAIA, xbench-deepsearch, WebWalkerQA, and Frames, the Tongyi DeepResearch model with 3B activated parameters outperformed ReAct Agents based on flagship models including OpenAI o3, DeepSeek V3.1, and Claude-4-Sonnet.