Tongyi DeepResearch is an open-source deep research agent from Alibaba’s Tongyi Lab, released on GitHub where it has collected more than 20,000 stars. It is an agentic large language model built specifically for long-horizon, deep information-seeking tasks: the kind of multi-step web research that produces cited, structured reports rather than one-shot answers. The model and agent code are fully open, and the project has been actively updated through 2026.
The released model, Tongyi-DeepResearch-30B-A3B, has 30.5 billion total parameters with only 3.3 billion activated per token, and a 128K context length. It was trained with a fully automated synthetic data pipeline, large-scale continual pre-training on agentic data, and end-to-end reinforcement learning with a customized GRPO framework. Tongyi Lab reports state-of-the-art results on agentic search benchmarks including Humanity’s Last Exam, BrowseComp, BrowseComp-ZH, WebWalkerQA, and FRAMES. At inference time the agent supports both a ReAct paradigm and a heavier IterResearch mode that uses test-time scaling for maximum performance.
The model weights and agent code are free and open source under the Apache-2.0 license. Hosted access is also available through API providers such as OpenRouter on a pay-per-use basis.
Tongyi DeepResearch is for researchers and developers building deep research agents, and for teams that want an open, self-hostable alternative to proprietary deep research products. It is also a useful reference implementation for anyone studying how agentic models are trained, from synthetic data generation to reinforcement learning at scale.