Tongyi DeepResearch: Open-Source Web Agent with Performance Comparable to OpenAI DeepResearch

In published benchmarks, it recorded 32.9 points on HLE, 43.4 points on BrowseComp, and 46.7 points on BrowseComp-ZH. On the user-centric xbench-DeepSearch, it scored 75 points, which is reported to outperform existing proprietary and open-source agents.

In addition to the model release, proven methodologies for creating advanced agents have been shared. This includes new data synthesis solutions across the entire training pipeline, from Agentic Continual Pre-training (CPT) to Reinforcement Learning (RL). For the RL stage, a full-stack solution is provided, encompassing algorithmic innovation, automated data curation, and robust infrastructure.

During inference, the model leverages its inherent capabilities without the need for prompt engineering via the vanilla ReAct framework. Furthermore, a "Heavy Mode" utilizing test-time scaling demonstrates the limits of complex reasoning and planning.

For data synthesis, the team proposes AgentFounder, a systematic and scalable solution. This forms a data flywheel using data from the post-training pipeline. Data is collected from diverse sources—such as documents, crawling data, and knowledge graphs—and restructured into an entity-anchored open-world knowledge memory.

This process generates large-scale synthetic QA pairs. In advanced action synthesis, historical trajectories are remodeled as multi-step decision-making processes, enabling the exploration of the inference action space in an offline environment. This eliminates the need for additional commercial tool API calls.