explaingit

alibaba-nlp/deepresearch

★ 18,891PythonAudience · developerComplexity · 4/5MaintainedLicenseSetup · hard

TLDR

An AI research agent that automatically searches the web, reads multiple sources, and synthesizes findings to answer complex research questions end-to-end.

Mindmap

mindmap
  root((repo))
    What it does
      Web search automation
      Multi-source synthesis
      Complex reasoning
    How it works
      30.5B parameters
      ReAct mode
      IterResearch mode
    Use cases
      Automated research
      AI assistant building
      Agent benchmarking
    Tech stack
      Python
      HuggingFace
      OpenAI API
    Getting started
      Download weights
      Online demo
      Third-party APIs

Things people build with this

USE CASE 1

Build an AI assistant that investigates complex questions by searching and reading web sources automatically.

USE CASE 2

Automate research workflows for topics requiring synthesis of information from multiple documents.

USE CASE 3

Benchmark and evaluate AI research agent performance on multi-step information-seeking tasks.

Tech stack

PythonHuggingFaceModelScopeOpenAI APISerperJina

Getting it running

Difficulty · hard Time to first run · 1h+

Requires API keys for OpenAI and Serper, plus potential HuggingFace/ModelScope authentication and model downloads.

Use freely for any purpose including commercial. Keep the notice and disclose changes to the patent grant.

In plain English

Tongyi DeepResearch is an open-source AI research agent developed by Alibaba's Tongyi Lab. It is designed to tackle complex, multi-step research questions that require searching the web, reading multiple sources, and synthesizing findings, much like a human researcher would, but automated. The underlying model has 30.5 billion parameters total, with 3.3 billion activated per question, giving it a balance of capability and efficiency. The model specializes in "deep information-seeking" tasks: questions that can't be answered with a single search result but require browsing, reading, comparing, and reasoning across many web pages. It supports two operating modes, a lightweight ReAct mode for direct evaluation, and a heavier IterResearch mode that uses extra compute at inference time to push for higher-quality answers. You would use it when you need automated research on complex topics, want to build an AI assistant that can investigate questions end-to-end, or are benchmarking AI research agents. Researchers can download the model weights from HuggingFace or ModelScope, or try it via an online demo. Running it requires Python 3.10, and it integrates with third-party APIs for web search (Serper), web page reading (Jina), and an OpenAI-compatible language model for summarization.

Copy-paste prompts

Prompt 1
How do I download and run Tongyi DeepResearch locally to answer research questions?
Prompt 2
Show me how to integrate Tongyi DeepResearch with my own web search API and summarization model.
Prompt 3
What's the difference between ReAct mode and IterResearch mode, and when should I use each?
Open on GitHub → Explain another repo

Generated 2026-05-18 · Model: sonnet-4-6 · Verify against the repo before relying on details.