Alibaba-NLP/DeepResearch presents Tongyi DeepResearch, a 30B-parameter agent-focused LLM with on-policy RL, designed for long-horizon information seeking and agentic search tasks. It provides a Quick Start guide, environment setup, and inference options via local scripts or OpenRouter.
Collecting history — the radar snapshots this repo daily. The trend line appears after 3 days of data (1 so far).
What it is
Tongyi DeepResearch is an agentic large language model described as having 30.5 billion total parameters with 3.3 billion activated per token, developed for long-horizon, deep information-seeking tasks. It supports two inference paradigms at test time: ReAct and an IterResearch-based 'Heavy' mode. The project ships with a model download option and benchmark materials.
How it works
The project emphasizes an end-to-end workflow for data synthesis, continual pre-training on agentic data, and on-policy reinforcement learning using a customized Group Relative Policy Optimization framework with token-level policy gradients and selective filtering of negative samples. It also describes compatibility with multiple inference paradigms (ReAct and Heavy mode) for evaluation.
Getting started
- Environment setup guidance is provided with: "Recommended Python version: 3.10.0 (using other versions may cause dependency issues)." and sample Conda commands.
- Installation: "```bash pip install -r requirements.txt
- Configuration: copy and edit the environment file: "cp .env.example .env" and populate keys such as SERPER_KEY_ID, JINA_API_KEYS, API_KEY/API_BASE, DASHSCOPE_API_KEY, SANDBOX_FUSION_ENDPOINT, MODEL_PATH, DATASET, OUTPUT_PATH.
- Data formats: supports JSON and JSONL evaluation data; JSONL is recommended with fields "question" and "answer".
- Inference scripts: instructions to modify run_react_infer.sh variables including MODEL_PATH, DATASET, OUTPUT_PATH, and required API keys.
- Inference execution: "bash run_react_infer.sh".
## Recent releases
- RELEASES (latest 0): none
## Traction
- Stars: 19783
- Forks: 1504
## Behind the repo
- Organization: Alibaba-NLP; links to HuggingFace and ModelScope model pages are provided in the README, along with blog, paper, and demo references.
## Caveats
- License: Apache-2.0
- Creation date: 2025-01-09; last push: 2026-02-27
- Open issues: 92
- Language: Python
- Notes indicate that the online demo has latency and QPS limits; for stability, local deployment is recommended.