TLDR: Alibaba Cloud’s Tongyi Lab has open-sourced its groundbreaking Tongyi DeepResearch Agent, a 30-billion-parameter Mixture of Experts (MoE) model designed for complex, multi-step information retrieval. This lightweight AI agent has demonstrated state-of-the-art performance across various benchmarks, surpassing models from OpenAI and DeepSeek, and is now available to developers on platforms like GitHub and Hugging Face.
Alibaba Cloud’s Tongyi Lab has announced the full open-sourcing of its innovative Tongyi DeepResearch Agent, marking a significant milestone in the development of AI agents. Released on September 17, 2025, this web proxy tool is engineered for long-horizon, deep information-seeking tasks, offering a powerful foundation for developers and small teams to build specialized AI agents.
At its core, Tongyi DeepResearch is a 30-billion-parameter Mixture of Experts (MoE) model, uniquely designed to activate only 3 billion parameters per token. This sparse activation mechanism ensures high efficiency and robust performance, even on resource-constrained hardware, while maintaining an ultra-long context window of 128K tokens. This capability is crucial for handling extensive document chains and multi-threaded web searches, enabling the agent to efficiently decompose complex questions, perform multi-step reasoning, and integrate massive datasets.
The agent has achieved state-of-the-art (SOTA) results across several authoritative benchmark datasets, demonstrating its superior capabilities. In Humanity’s Last Exam (HLE), Tongyi DeepResearch scored 32.9 points, outperforming OpenAI’s o3 model (24.9 points). It also achieved 75.0 points in xbench-DeepSearch, surpassing o3’s 67.0 points. Against Google’s Gemini, it claimed a 35.2 score on BrowseComp-ZH, a 10-point lead. Furthermore, its performance on FRAMES, a tool-use benchmark, reached 90.6, dwarfing Claude 3.5 Sonnet’s 84.3, and significantly outperforming other open-source peers like Llama variants (e.g., 21.1 on HLE).
Key technological innovations underpinning Tongyi DeepResearch include continuous pre-training on agentic data and an on-policy reinforcement learning approach. These advancements contribute to its strong stability and reliability in diverse scenarios, such as multi-day itinerary planning and legal document analysis. The agent offers two primary reasoning modes: a single-model ReAct mode, suitable for lightweight tasks, and a ‘Heavy’ mode that supports parallel collaboration of multiple agents and scalable computing power.
Also Read:
- Sentient AI Unveils Open AGI Network, Challenging Tech Giants with Collaborative Intelligence
- OpenAI Envisions a Future of Cloud-Based Coding Powered by Millions of AI Agents
The open-sourcing of Tongyi DeepResearch is expected to significantly lower the development barrier for AI agents, providing a powerful, accessible tool for innovation. The model, its framework, and solutions are now fully available for download from platforms like GitHub, Hugging Face, and ModelScope communities, fostering broader adoption and further advancements in the AI agent ecosystem. This release follows Alibaba Cloud’s previous efforts, which have seen over 200 models of varying parameter scales open-sourced, leading to more than 100,000 secondary developed models on Hugging Face alone.
