Tongyi DeepResearch 30B-A3B

Tongyi DeepResearch 30B-A3B, published on 16 September 2025, is not a chat model with a search button. It is trained end to end for long-horizon information seeking: issue a query, read results, issue the next query, and keep going for dozens of steps before writing an answer. Alibaba positions it as the first fully open-weights web agent that matches OpenAI's Deep Research, a service that is not available for download at any price. The architecture is a mixture of experts: 48 layers, hidden dimension 2,048, 128 experts of which 8 fire per token, giving 30 billion parameters in storage and about 3 billion in use at any moment. The window is 131,072 tokens — a practical requirement rather than a boast, since a search agent accumulates page after page of retrieved text inside a single run. It runs in two modes: plain ReAct, which the vendor uses to measure the model's own ability, and a heavier iterative mode that spends more computation at answer time to push the ceiling higher. Vendor scores: 32.9 on Humanity's Last Exam, 43.4 on BrowseComp and 46.7 on its Chinese counterpart BrowseComp-ZH, and 75 on xbench-DeepSearch. As always with agent benchmarks, the number depends on the scaffolding around the model as much as the weights, and these are the vendor's own runs. The most interesting admission is buried in the engineering notes. Alibaba trained the model chiefly on synthetic tasks generated by its own pipeline, and reports that training directly on the BrowseComp test set produced substantially worse results than training on the synthetic data. Their explanation is that human-annotated agent data is noisy and too small to describe a learnable distribution. It is a rare case of a vendor publishing a negative result about the obvious shortcut. Weights are Apache 2.0, with inference scripts in a separate GitHub repository. Adoption in the 30 days to 23 September 2026: 29,537 downloads against 821 likes — a ratio that marks it as a model far more admired than deployed, which fits a 30-billion-parameter agent that needs a search stack around it before it does anything at all.

▸Videos

No videos yet.

▸Reviews