Appearance
Tongyi DeepResearch
Category: OSS Agent
Official GitHub: https://github.com/Alibaba-NLP/DeepResearch
Blog: https://tongyi-agent.github.io/blog/introducing-tongyi-deep-research/
Paper: https://arxiv.org/pdf/2510.24701
License: Apache 2.0 (open-source, free to self-host)
Pricing: Free and open-source. You can also access the model via OpenRouter (pay-per-token, no GPU required) or through Alibaba Cloud's Bailian service (managed, paid).
What it is
Tongyi DeepResearch is an agentic large language model (30.5 billion total parameters, 3.3 billion activated per token, Mixture-of-Experts architecture) developed by Tongyi Lab at Alibaba. It is purpose-built for long-horizon, deep information-seeking tasks -- the class of work where an agent must plan, search the web, read documents, reason across sources, and synthesize a thorough answer over many steps.
It achieves state-of-the-art results on agentic search benchmarks including Humanity's Last Exam, BrowseComp, BrowseComp-ZH, WebWalkerQA, xbench-DeepSearch, FRAMES, and SimpleQA.
Key capabilities
- ReAct inference mode -- rigorous step-by-step tool use (search, read, reason, answer) for evaluating core model abilities.
- IterResearch (Heavy) mode -- test-time scaling that trades more compute for higher accuracy on hard research questions.
- 128K context window -- can hold long research threads, large document sets, or multi-turn conversations in context.
- Tool-augmented -- integrates web search (Serper), page reading (Jina), a Python interpreter sandbox (SandboxFusion), and file parsing (Dashscope) at inference time.
Model download
| Model | HuggingFace | ModelScope | Size | Context |
|---|---|---|---|---|
| Tongyi-DeepResearch-30B-A3B | HuggingFace | ModelScope | 30B-A3B | 128K |
Try it without a GPU
- ModelScope online demo
- Hugging Face Space
- OpenRouter API (pay-per-token)
Note: The online demos are for quick exploration. Response times may vary due to model latency and tool QPS limits. For stable production use, run locally or use the Bailian managed service.
Safety and ToS notes
- The model weights are released under the Apache 2.0 license. Commercial use is permitted subject to that license.
- The inference pipeline calls third-party APIs (Serper, Jina, OpenAI-compatible endpoint, Dashscope). Each provider's own terms apply to those calls.
- The optional Python sandbox (SandboxFusion) executes arbitrary code. Run it in an isolated environment; do not expose it to untrusted inputs without proper sandboxing.
- The
.envfile containing API keys is gitignored by default; do not commit it.