In the wee hours of September 17, Alibaba made its first deep research Agent model, Tongyi DeepResearch, available as open-source. This model delivered state-of-the-art (SOTA) performance on several authoritative benchmarks, including HLE, BrowseComp-zh, and GAIA, outperforming models like OpenAI's Deep Research and DeepSeek-V3.1. Presently, the model, along with its framework and solution, is fully open-sourced, enabling users to download both the model and its code from the Github, Hugging Face, and ModelScope communities.

To tackle the challenges of "cognitive overload" and "irreversible noise interference" that current deep research models encounter when dealing with long-term tasks, the Tongyi team has developed a comprehensive training pipeline powered by synthetic data. This pipeline covers both pre-training and post-training stages, substantially boosting the model's iteration speed and generalization abilities. During the inference stage, the team introduced two modes: ReAct and Heavy (built on their proprietary IterResearch), facilitating high-quality reasoning even in extended tasks.