Tongyi has released Tongyi DeepResearch, the first fully open-source web agent to achieve performance comparable to OpenAI’s DeepResearch across a comprehensive suite of benchmarks. The model scores 32.9 on Humanity’s Last Exam (HLE), 43.4 on BrowseComp, and 75 on xbench-DeepSearch, systematically outperforming existing proprietary and open-source alternatives.
The release includes a complete methodology for training advanced agents, featuring Agentic Continual Pre-training (CPT) via the AgentFounder data synthesis solution. The team details a novel pipeline for generating high-quality synthetic QA pairs using graph-based methods and iterative complexity upgrades to create PhD-level research questions. Inference supports both native ReAct Mode for intrinsic capability benchmarking and Heavy Mode (test-time-scaling) via the IterResearch paradigm for complex multi-step tasks.
This contribution provides a battle-tested framework for creating deep research agents, offering algorithmic innovations and automated data curation that eliminate the need for commercial tool API calls during training.