Tongyi DeepResearch
github.com/alibaba-nlp/deepresearch- Category
- AI Agents
- Rank
- No. 90Tools index
- Pricing
- Open Source
- Type
- AGENT
- Builder
- alibaba-nlp
- GitHub
- 19.9k stars
- Date
About
An open-weights agentic model from Alibaba's Tongyi Lab built for long-horizon information seeking: it plans, searches and reads across many steps to answer research questions a single prompt cannot. It uses 30.5B total parameters with 3.3B activated per token, and runs either in a plain ReAct loop or in a "Heavy" mode that spends extra test-time compute. The repo also documents the training recipe behind it, from synthetic agentic data generation through continual pre-training to end-to-end reinforcement learning.
What it does
Tongyi DeepResearch runs an iterative research loop that chooses among web search, academic search, page reading, file parsing, and sandboxed Python. It gathers tool responses until it produces a tagged final answer or reaches configured call, time, or context limits.
Why it's ranked here
The tool offers an unusually complete research harness, not just model weights. It combines source discovery, page extraction, document handling, computation, parallel rollouts, and benchmark judges. That breadth is valuable, but local operation demands substantial infrastructure, credentials, and configuration.
What's good
Its research actions cover complementary evidence channels. Search accepts batches, academic lookup handles multiple queries concurrently, page visits extract goal-specific evidence, and the parser supports documents, spreadsheets, archives, audio, and video. Runs preserve questions, reference answers, conversations, predictions, and termination reasons for later evaluation.
Tradeoffs
Setup is heavy. The pinned dependency list includes CUDA libraries, PyTorch, vLLM, Ray, document tooling, and many cloud clients. Full operation can require separate credentials for search, page reading, summarization, file parsing, and Python sandboxes. The hosted demos may also fail intermittently because of latency and tool limits.
How to use it well
Use it for controlled research experiments, benchmark runs, and questions that benefit from repeated searching, source reading, document inspection, or computation. Prepare JSON or JSONL questions, keep reference answers when evaluating, and run isolated environments. It does not provide a lightweight desktop research interface or a self-contained offline workflow.
Technical notes+
inference/react_agent.py implements MultiTurnReactAgent, registers five tools, targets an OpenAI-compatible service on localhost, permits up to 100 model calls, stops after 150 minutes, and forces an answer when conversation tokens exceed 110 KiB. inference/tool_search.py and inference/tool_scholar.py use Serper endpoints; inference/tool_visit.py reads through Jina and summarizes through an OpenAI-compatible API; inference/tool_python.py dispatches code to SandboxFusion; inference/tool_file.py routes documents through Qwen Agent parsers. inference/run_multi_react.py reads JSON or JSONL, supports dataset splitting and parallel rollouts, and writes JSONL results. evaluation/evaluate_hle_official.py uses a structured LLM judge. requirements.txt pins the full Python and CUDA-oriented environment.
Observed
- Primary language
- Python
- Installation surface
- Pinned dependencies installed from requirements.txt, with Python 3.10 recommended
- Execution interface
- Command-line inference through a shell script and Python runner
- Input formats
- JSON arrays and JSONL records containing question and answer fields
- Model access
- Downloadable through Hugging Face or ModelScope, with OpenRouter API access documented
- Research tools
- Web search, Google Scholar search, webpage visiting, file parsing, and sandboxed Python
- Acceleration stack
- Pinned CUDA 12, PyTorch, vLLM, Triton, and xFormers dependencies
Read from README.md, requirements.txt, inference/prompt.py, evaluation/prompt.py, inference/tool_file.py, inference/tool_visit.py, inference/react_agent.py, inference/tool_python.py, inference/tool_search.py, inference/tool_scholar.py, inference/run_multi_react.py, evaluation/evaluate_hle_official.py.
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.