Tongyi DeepResearch
StaleDescription
Open-source deep research agent from Alibaba Tongyi Lab, using multi-stage iterative information retrieval and reasoning to conduct deep analysis, synthesis, and summarization of complex topics with web search and document analysis.
Key Features
- 30.5B total parameter agentic LLM with only 3.3B activated per token (MoE architecture)
- Fully automated synthetic data generation pipeline for agentic pre-training, SFT, and RL
- Large-scale continual pre-training on diverse agentic interaction data
- End-to-end reinforcement learning with custom Group Relative Policy Optimization
- Compatible with ReAct and IterResearch-based Heavy inference paradigms
- State-of-the-art on Humanity's Last Exam, BrowseComp, FRAMES, and SimpleQA benchmarks
Use Cases
💡 Deep information-seeking and multi-step web research tasks
💡 Long-horizon agentic search and question answering
💡 Academic research and literature review automation
💡 Complex fact-checking across multiple web sources
💡 Deploying as a research assistant API service
Strengths & Limitations
✅ Strengths
- • High community interest (19.9k stars)
- • Permissive open-source license (Apache-2.0)
- • Responsive to issues, low backlog
⚠️ Limitations
- • No updates in over 6 months
Categories
Quick Start
1. Setup: conda create -n deepresearch python=3.10.0 && conda activate deepresearch
2. Install: pip install -r requirements.txt
3. Configure: cp .env.example .env and fill API keys (SERPER_KEY_ID, JINA_API_KEYS, etc.)
4. Run inference: python inference/react_inference.py --dataset <your_dataset>