Generates 96,000 high-quality deep research trajectories and provides a fully open-source recipe for training agentic language models on long-horizon web research tasks without external API dependencies.

https://t.co/zUo5XOpZBb https://t.co/HmfzWXxrHO

TIGER-AI-Lab/OpenResearcher

Source: https://github.com/TIGER-AI-Lab/OpenResearcher

🤗 HuggingFace ｜ Slack | WeChat

📣 News

• [2026.6.05] 🔥 OpenResearcher data has been adopted by Nemotron 3 Ultra!
• [2026.3.25] 🚀 OpenResearcher hits #2 on Hugging Face Daily Papers!
• [2026.3.24] OpenResearcher paper is now available, highlighting practical insights into deep research pipeline design. Already adopted by NVIDIA’s Nemotron family of models!
• [2026.2.25] Honored to be among the top 3 trending datasets on 🤗 Hugging Face — now 11K+ downloads!
• [2026.2.18] The OpenResearcher training code is now available. Start training your own OpenResearcher!
• [2026.2.14] Excited to have our OpenResearcher demo video. Dive in and unlock the power of Deep Research today!
• [2026.2.12] Excited to see OpenResearcher powering deep research trajectory generation in NVIDIA’s NeMo Data Designer!
• [2026.2.10] Our X post received 1.2K+ likes! Feel free to check out the post and join the discussion!

💥 Introduction

OpenResearcher is a fully open agentic large language model (30B-A3B) designed for long-horizon deep research scenarios. It achieves an impressive 54.8% accuracy on BrowseComp-Plus, surpassing performance of GPT-4.1, Claude-Opus-4, Gemini-2.5-Pro, DeepSeek-R1 and Tongyi-DeepResearch. We fully open-source the training and evaluation recipe—including data, model, training methodology, and evaluation framework for everyone to progress deep research.

🏆 Deep Research Benchmark Results

✨ Features

• 🔑 Fully Open-Source Recipe — We fully open-source our 96K high-quality DeepResearch trajectory dataset with 100+ turns generated by GPT-OSS-120B with native browser tools, the leading 30B-A3B model trained on it, distillation recipe, and a lightweight DeepResearch evaluation framework to progress deep research.

• 💰 Highly Scalable and Low-Cost — We generate DeepResearch trajectories at massive scale using self-built retriever over a dedicated ~11B-token corpus, eliminating the need for external Search APIs. This scalable retriever significantly reduces training costs.

• 🚀 Remarkable Performance on Deep Research Benchmarks — OpenResearcher demonstrates leading performance across a range of deep research benchmarks, including BrowseComp-Plus, BrowseComp, GAIA, xbench-DeepSearch.

📋 Table of Contents

• 🛠 Environment Setup
  • Installation
  • Deep Research Benchmarks Preparation
• 🔍 Configuration
• 🚀 Quick Start
• 🔬 Benchmark OpenResearcher
  • Example 1: BrowseComp-Plus with Local Search Engine
  • Example 2: GAIA with Serper API (No Local Search Needed)
  • Evaluation
  • Quick Commands
• 🧪 (Optional) Train Your Own OpenResearcher
• 🤝 Core Contributors
• 🎓 Advisors
• 🙏 Acknowledgements
• ✨ Contributing
• 📚 Citation

🛠 Environment Setup

We run this repo on the following setup:

• 8 * A100 80G Nvidia GPUs
• Linux operating system

Other hardware setups can also work, but remember to modify the corresponding parameters.

Installation

sudo apt update 
sudo apt install -y openjdk-21-jdk

# install uv
curl -LsSf https://astral.sh/uv/install.sh | sh
uv venv --python 3.12
source .venv/bin/activate

# install tevatron for BrowseComp-plus 
git clone https://github.com/texttron/tevatron.git
cd tevatron
uv pip install -e .
cd ..

# install all dependencies automatically
uv pip install -e .

Deep Research Benchmarks Preparation

Run the setup script to automatically download the BrowseComp-Plus benchmark. Other benchmarks, including BrowseComp, GAIA and xbench-DeepResearch, will be set up automatically when they are first used.

bash setup.sh

This script will:

• ✅ Verify Python 3.12 virtual environment and automatically install any missing dependencies
• ✅ Downlaod BrowseComp-Plus dataset from HuggingFace and set up the directory structure

For more info about these deep research benchmarks, see benchmarks.md

🔍 Configuration

Copy the template and configure your API keys:

cp .env.template .env

Edit .env:

# Serper API (for web search when using browser_backend=serper)
SERPER_API_KEY=your_key        # Get from: https://serper.dev/

# OpenAI API (for evaluation scoring)
OPENAI_API_KEY=your_key        # Get from: https://platform.openai.com/api-keys

🚀 Quick Start

Prerequisites: Install dependencies and configure API keys (see Environment Setup and Configuration)

1. Deploy OpenResearcher-30B-A3B:

bash scripts/start_nemotron_servers.sh

The complete vLLM server logs can be found in the logs directory.

2. Run your first task (Before proceeding, check the logs in logs directory to ensure the vLLM server is deployed.)

import asyncio
from deploy_agent import run_one, BrowserPool
from utils.openai_generator import OpenAIAsyncGenerator

async def main():
    # Initialize generator and browser
    generator = OpenAIAsyncGenerator(
        base_url="http://localhost:8001/v1",
        model_name="OpenResearcher/OpenResearcher-30B-A3B",
        use_native_tools=True
    )
    browser_pool = BrowserPool(search_url=None, browser_backend="serper")

    # Run deep research
    await run_one(
        question="What is the latest news about OpenAI?",
        qid="quick_start",
        generator=generator,
        browser_pool=browser_pool,
    )

    browser_pool.cleanup("quick_start")

if __name__ == "__main__":
    asyncio.run(main())

The deep research agent will automatically search the web, browse webpages, and extract relevant information. You’ll see the final answer along with all intermediate reasoning steps.

🔬 Benchmark OpenResearcher

We benchmark our OpenResearcher-30B-A3B using below deep research benchmarks:

For more info about these deep research benchmarks, see benchmarks.md

Example 1: BrowseComp-Plus with Local Search Engine

Complete evaluation using local dense search with browsecomp-plus corpus and embeddings (note: only applicable for BrowseComp-Plus):

# Terminal 1: Start local Dense search service on port 8000
# Embedding model (Qwen3-Embedding-8B) will be deployed on GPUs 7
bash scripts/start_search_service.sh dense 8000

# Terminal 2: Start vLLM servers (requires 4 GPUs)
# TP=2, deploy 2 servers starting from port 8001 on GPUs 0,1,2,3
bash scripts/start_nemotron_servers.sh 2 8001 0,1,2,3

# Terminal 3: Run agent
bash run_agent.sh results/browsecomp_plus/OpenResearcher_dense 8001 2 browsecomp_plus local OpenResearcher/OpenResearcher-30B-A3B

What this does:

• Deploys Dense retriever service on port 8000 as search engine
• Launches 2 vLLM servers (ports 8001, 8002) with TP=2 across 4 GPUs
• Runs deepresearch agent with load balancing across both servers

Example 2: GAIA with Serper API (No Local Search Needed)

Run with Serper Google Search API (note: applicable to all benchmarks except BrowseComp-Plus):

# Terminal 1: Start vLLM servers (requires 4 GPUs)
bash scripts/start_nemotron_servers.sh 2 8001 0,1,2,3

# Terminal 2: Run agent with serper search backend
bash run_agent.sh results/gaia/OpenResearcher_serper 8001 2 gaia serper OpenResearcher/OpenResearcher-30B-A3B

Browser Backend Options:

• local - Use local BM25/Dense search service (for BrowseComp-Plus)
• serper - Use Serper Google Search API (for all other benchmarks)

For other parameters, refer to parameter.md.

Evaluation

After running experiments, evaluate results:

# eval on browsecomp_plus
python eval.py --input_dir results/browsecomp_plus_dense/OpenResearcher_dense

# eval on gaia
python eval.py --input_dir results/gaia/OpenResearcher_serper

Quick Commands

For script parameter explanation, refer to parameter.md.

Note: Don’t forget to evaluate your results using:

python eval.py --input_dir [INPUT_DIR]

🧪 (Optional) Train Your Own OpenResearcher

Our OpenResearcher-30B-A3B is trained using Megatron-LM on openresearcher-dataset. To get started, clone the openresearcher branch of the Megatron-LM repository:

git clone -b openresearcher https://github.com/jdf-prog/Megatron-LM.git

Then, follow the training instructions here to train your own OpenResearcher!

🤝 Core Contributors

Zhuofeng Li

Dongfu Jiang

Xueguang Ma

Haoxiang Zhang

Ping Nie

🎓 Advisors

Wenhu Chen

Yu Zhang

🙏 Acknowledgements

✨ Contributing

We are truly looking forward to open-source contributions to OpenResearcher! If you’re interested in contributing, collaborating, or reporting issues, please feel free to open an issue or submit a pull request (PR). You can also reach us at [email protected].

We are also looking forward to your feedback and suggestions!

📚 Citation

@article{li2026openresearcher,
  title={{OpenResearcher: A Fully Open Pipeline for Long-Horizon Deep Research Trajectory Synthesis}},
  author={Li, Zhuofeng and Jiang, Dongfu and Ma, Xueguang and Zhang, Haoxiang and Nie, Ping and Zhang, Yuyu and Zou, Kai and Xie, Jianwen and Zhang, Yu and Chen, Wenhu},
  journal={arXiv preprint arXiv:2603.20278},
  year={2026}
}

⭐ Star History

↑ Back to Top ↑