Tongyi DeepResearch Agent

A Tongyi DeepResearch Agent is an open-source AI agent that conducts autonomous multi-step web research and compiles comprehensive reports using a 30B parameter model architecture derived from Qwen3-30B-A3B.

• AKA: Tongyi Deep Research Agent, DeepResearch Agent, DeepResearch, Alibaba Tongyi DeepResearch System, Tongyi Research Agent.
• Context:
  • It can typically perform Tongyi DeepResearch Web Research Tasks with web traversal tools and search APIs.
  • It can typically generate Tongyi DeepResearch Research Reports through synthetic data training and reinforcement learning optimization.
  • It can typically operate in ReAct Mode for benchmark evaluation tasks.
  • It can typically operate in Heavy Mode for complex research synthesis tasks.
  • It can often achieve Competitive Performance on Humanity's Last Exam (HLE) Benchmarks scoring above 70%.
  • It can often integrate with Alibaba Platform Services through API interfaces.
  • It can often support Real-World Deployments in travel planning domains and legal research domains.
  • It can range from being a Simple Tongyi DeepResearch Agent to being a Complex Tongyi DeepResearch Agent, depending on its task complexity level.
  • It can range from being a Single-Query Tongyi DeepResearch Agent to being a Multi-Step Tongyi DeepResearch Agent, depending on its research depth requirement.
  • It can range from being a Baseline Tongyi DeepResearch Agent to being a Fine-Tuned Tongyi DeepResearch Agent, depending on its training iteration count.
  • It can range from being a ReAct Mode Tongyi DeepResearch Agent to being a Heavy Mode Tongyi DeepResearch Agent, depending on its inference complexity setting.
  • ...
• Example(s):
  • Tongyi DeepResearch Agent Deployments, such as:
    • Gaode Mate Integration, which uses the agent for travel itinerary generation.
    • Tongyi FaRui Implementation, which applies the agent to legal case retrieval.
  • Tongyi DeepResearch Agent Components, such as:
    • AgentFounder-30B Model, which provides agentic behavior pretraining.
    • WebSailor-V2-30B-A3B Model, which enables web environment navigation.
  • ...
• Counter-Example(s):
  • Proprietary Research Agent, which lacks open source availability.
  • Single-Step Query System, which cannot perform multi-step reasoning.
  • Static Knowledge System, which lacks web traversal capability.
• See: Open-Source AI Agent, DeepSeek Model Family, Alibaba Tongyi Lab, Web Research Agent, Chinese AI Agent Ecosystem, AI Agent Architecture, Deep Research System, Qwen Model Series, Reinforcement Learning Agent, Apache 2.0 License.

---