DR-Venus 是一个4B 参数的深度研究智能体，完全基于开放数据训练而成。它在多个深度研究基准上确立了小模型的新前沿，证明在边缘规模下，通过精心的数据筛选和有效的训练策略，可以涌现出强大的智能体能力。
图 1：DR-Venus 训练流程概览。
| 骨干模型 | Qwen3-4B-Thinking-2507 |
| 训练数据 | 仅开放数据（REDSearcher Data） |
| 工具协议 | search + visit |
| 交互步数上限 | 最多 200 个工具调用步骤 |
| 上下文长度 | 200K+（训练）/ 256K（推理） |
🔥 新闻
2026-04-24发布了 GGUF 版本的模型 checkpoint：DR-Venus-4B-SFT-GGUF和DR-Venus-4B-RL-GGUF。2026-04-23我们的技术报告现已在 arXiv 和 HF Daily Paper 上发布（当日论文第 3 名）。2026-04-22发布了模型 checkpoint：DR-Venus-4B-SFT和DR-Venus-4B-RL。2026-04-22开源了完整的训练和推理代码库。
📖 概述
DR-Venus 的核心目标是在有限的开放数据监督下，通过提升数据质量和有效数据利用率，构建一个强大的边缘规模深度研究智能体。该项目包含三个阶段：
| 阶段 | 描述 | 代码 |
|---|---|---|
| 1. SFT | 将原始的 REDSearcher 轨迹转换为统一的智能体格式，清理嘈杂的工具交互，筛选出正确的内容，并在监督微调之前通过轮次感知的重采样对长时程轨迹进行加权提升。 | SFT/ |
| 2. RL | 从 SFT 检查点出发，应用长时程强化学习，采用 IGPO 风格的信息增益奖励和轮次级别的格式感知惩罚。 | RL/ |
| 3. 推理 | 使用与训练时相同的 search + visit 工具协议部署训练好的模型。 | Inference/ |
📊 主要结果
DR-Venus-4B 在多个深度研究基准上确立了强大的小模型前沿水平。
与小规模开源模型的对比
| 模型 | BrowseComp | BrowseComp-ZH | GAIA（文本） | xBench-DS-2505 | xBench-DS-2510 | DeepSearchQA |
|---|---|---|---|---|---|---|
| DeepDive-9B-SFT | 5.6 | 15.7 | -- | 35.0 | -- | -- |
| DeepDive-9B-RL | 6.3 | 15.1 | -- | 38.0 | -- | -- |
| WebSailor-7B | 6.7 | 14.2 | 37.9 | 34.3 | -- | -- |
| OffSeeker-8B-SFT | 10.6 | 24.2 | 47.6 | 48.0 | -- | -- |
| OffSeeker-8B-DPO | 12.8 | 26.6 | 51.5 | 49.0 | -- | -- |
| WebExplorer-8B-RL | 15.7 | 32.0 | 50.0 | 53.7 | 23.0 | 17.8 |
| AgentCPM-Explore-4B | 24.1 | 29.1 | 63.9 | 70.0 | 34.0 | 32.8 |
| DR-Venus-4B-SFT | 26.8 | 35.7 | 65.4 | 69.0 | 35.3 | 37.7 |
| DR-Venus-4B-RL | 29.1 | 37.7 | 64.4 | 74.7 | 40.7 | 39.6 |
核心要点：
DR-Venus-4B-SFT在大多数被跟踪的基准上已经超越了此前的小型开源智能体。DR-Venus-4B-RL在 SFT 的基础上进一步提升，在 BrowseComp 上提升 +2.3，在 BrowseComp-ZH 上提升 +2.0。- RL 主要提升了长时程执行的可靠性、格式稳定性以及工具使用的校准能力。
图 2：在 BrowseComp 和 BrowseComp-ZH 上的 Pass@K 对比。
📦 数据流水线
SFT 流水线基于开放的 REDSearcher 轨迹构建，重点在于让有限的监督信号对小模型更有用。
Raw REDSearcher Trajectories (10,001)
│
├── Structural Cleaning ── environment alignment, tool normalization,
│ disallowed-tool pruning, duplicate removal
├── Correctness Filtering ── keep trajectories with correct final answers
│ → 9,365 trajectories
└── Turn-Aware Resampling ── upweight longer trajectories to emphasize
deep research behavior
→ 18,745 final SFT instances
在 RL 方面，我们使用从 REDSearcher RL 数据源 中整理出的开放 QA 监督数据。
✨ 演示
DR-Venus 演示视频
针对问题的 DeepResearch：“《真探》第一季里有一句类似‘useless spin’的台词出现在第几集？”。
🚀 快速开始
每个子目录都有自己的依赖项和入口脚本。完整细节请参阅各子项目的 README。
1. 推理
使用 search 和 visit 工具运行训练好的模型：
cd Inference
pip install -r requirements.txt
# Configure API credentials used by the tool server first
bash run_demo.sh
# Or launch the web demo
bash run_web_demo.sh --model_path <MODEL_PATH> --num_gpus <NUM_GPUS>
完整设置指南请参阅
Inference/README.md。
2. SFT
准备清洗后的轨迹并训练 SFT checkpoint：
cd SFT
pip install -e .
pip install -r requirements.txt
# Optional: convert raw RED trajectories into SFT-ready parquet
python data_clean/prepare_trajectories.py \
--input <RAW_PARQUET_OR_DIR> \
--output-dir <OUTPUT_DIR>
# Run supervised fine-tuning
bash train_sft.sh
数据格式、环境变量和 checkpoint 合并请参阅
SFT/README.md。
3. 强化学习
从 SFT 检查点继续，进行长时程强化学习：
cd RL
pip install -r requirements.txt
# Edit .env and train_igpo.sh first
bash train_igpo.sh
有关奖励配置、rollout 设置和故障排查，请参见
RL/README.md。
🔌 外部服务
Inference/ 和 RL/ 都依赖外部工具和模型端点。SFT 不需要这些。
| 服务 | 用途 |
|---|---|
| Serper | 网页搜索 |
| Jina Reader | 网页抓取 |
| OpenAI 兼容 API | 页面摘要 |
| OpenAI 兼容评判模型 | 强化学习奖励评估 |
📁 仓库结构
DR-Venus/
├── README.md
├── assets/
│ ├── dr-venus-overview.png
│ └── dr-venus-passk.png
├── SFT/
│ ├── data_clean/ # RED trajectory conversion & cleaning
│ ├── scripts/ # Checkpoint merge utilities
│ ├── sft_shells/
│ ├── train_sft.sh
│ └── README.md
├── RL/
│ ├── configs/
│ ├── data/
│ ├── tool_server/
│ ├── train_igpo.sh
│ └── README.md
├── Inference/
│ ├── tool_server/
│ ├── run_demo.sh
│ ├── run_web_demo.sh
│ └── README.md
└── model_cards/
🤝 致谢
DR-Venus 建立在多个强大的开源基础之上：
- verl —— 训练基础设施
- IGPO —— 强化学习算法基础
- Tongyi DeepResearch —— 深度研究智能体设计参考
- REDSearcher —— 开放数据轨迹及相关工具
📝 引用
@article{venus2026drvenus,
title={DR-Venus: Towards Frontier Edge-Scale Deep Research Agents with Only 10K Open Data},
author={Venus Team and Dai, Sunhao and Deng, Yong and Lin, Jinzhen and Song, Yusheng and Wang, Guoqing and Wu, Xiaofeng and Zhou, Yuqi and Yang, Shuo and Ying, Zhenzhe and Zhang, Zhanwei and Meng, Changhua and Wang, Weiqiang},
journal={arXiv preprint arXiv:2604.19859},
year={2026}
}
DR-Venus is a 4B-parameter deep research agent trained entirely on open data. It establishes a new small-model frontier on multiple deep research benchmarks, demonstrating that strong agentic capabilities can emerge from careful data curation and effective training strategies at edge scale.
Figure 1: Overview of the DR-Venus training pipeline.
| Backbone | Qwen3-4B-Thinking-2507 |
| Training Data | Open-data only (REDSearcher Data) |
| Tool Protocol | search + visit |
| Interaction Horizon | Up to 200 tool-call steps |
| Context Length | 200K+ (training) / 256K (inference) |
🔥 News
2026-04-24Released model checkpoints of GGUF version:DR-Venus-4B-SFT-GGUFandDR-Venus-4B-RL-GGUF.2026-04-23Our technical report is now available on arXiv and HF Daily Paper (#3 Paper of the day).2026-04-22Released model checkpoints:DR-Venus-4B-SFTandDR-Venus-4B-RL.2026-04-22Open-sourced the full training and inference codebase.
📖 Overview
The core goal of DR-Venus is to build a strong edge-scale deep research agent under limited open-data supervision by improving both data quality and effective data utilization. The project consists of three stages:
| Stage | Description | Code |
|---|---|---|
| 1. SFT | Convert raw REDSearcher trajectories into a unified agent format, clean noisy tool interactions, filter for correctness, and upweight long-horizon traces via turn-aware resampling before supervised fine-tuning. | SFT/ |
| 2. RL | Starting from the SFT checkpoint, apply long-horizon reinforcement learning with IGPO-style information gain rewards and turn-level format-aware penalties. | RL/ |
| 3. Inference | Deploy the trained model with the same search + visit tool protocol used during training. | Inference/ |
📊 Main Results
DR-Venus-4B establishes a strong small-model frontier on multiple deep research benchmarks.
Comparison with Small Open Models
| Model | BrowseComp | BrowseComp-ZH | GAIA (Text) | xBench-DS-2505 | xBench-DS-2510 | DeepSearchQA |
|---|---|---|---|---|---|---|
| DeepDive-9B-SFT | 5.6 | 15.7 | -- | 35.0 | -- | -- |
| DeepDive-9B-RL | 6.3 | 15.1 | -- | 38.0 | -- | -- |
| WebSailor-7B | 6.7 | 14.2 | 37.9 | 34.3 | -- | -- |
| OffSeeker-8B-SFT | 10.6 | 24.2 | 47.6 | 48.0 | -- | -- |
| OffSeeker-8B-DPO | 12.8 | 26.6 | 51.5 | 49.0 | -- | -- |
| WebExplorer-8B-RL | 15.7 | 32.0 | 50.0 | 53.7 | 23.0 | 17.8 |
| AgentCPM-Explore-4B | 24.1 | 29.1 | 63.9 | 70.0 | 34.0 | 32.8 |
| DR-Venus-4B-SFT | 26.8 | 35.7 | 65.4 | 69.0 | 35.3 | 37.7 |
| DR-Venus-4B-RL | 29.1 | 37.7 | 64.4 | 74.7 | 40.7 | 39.6 |
Key takeaways:
DR-Venus-4B-SFTalready outperforms prior small open agents on most tracked benchmarks.DR-Venus-4B-RLfurther improves over SFT by +2.3 on BrowseComp and +2.0 on BrowseComp-ZH.- RL mainly improves long-horizon execution reliability, formatting stability, and tool-use calibration.
Figure 2: Pass@K comparison on BrowseComp and BrowseComp-ZH.
📦 Data Pipeline
The SFT pipeline is built from open REDSearcher trajectories and focuses on making limited supervision more useful for a small model.
Raw REDSearcher Trajectories (10,001)
│
├── Structural Cleaning ── environment alignment, tool normalization,
│ disallowed-tool pruning, duplicate removal
├── Correctness Filtering ── keep trajectories with correct final answers
│ → 9,365 trajectories
└── Turn-Aware Resampling ── upweight longer trajectories to emphasize
deep research behavior
→ 18,745 final SFT instances
For RL, we use open QA supervision curated from the REDSearcher RL data source.
✨ Demo
DR-Venus Demo Video
DeepResearch on problem:"《真探》第一季里有一句类似“useless spin”的台词出现在第几集？".
🚀 Quick Start
Each subdirectory has its own dependencies and entry scripts. Refer to the subproject READMEs for full details.
1. Inference
Run the trained model with search and visit tools:
cd Inference
pip install -r requirements.txt
# Configure API credentials used by the tool server first
bash run_demo.sh
# Or launch the web demo
bash run_web_demo.sh --model_path <MODEL_PATH> --num_gpus <NUM_GPUS>
See
Inference/README.mdfor the full setup guide.
2. SFT
Prepare cleaned trajectories and train the SFT checkpoint:
cd SFT
pip install -e .
pip install -r requirements.txt
# Optional: convert raw RED trajectories into SFT-ready parquet
python data_clean/prepare_trajectories.py \
--input <RAW_PARQUET_OR_DIR> \
--output-dir <OUTPUT_DIR>
# Run supervised fine-tuning
bash train_sft.sh
See
SFT/README.mdfor data format, environment variables, and checkpoint merging.
3. RL
Continue from the SFT checkpoint with long-horizon RL:
cd RL
pip install -r requirements.txt
# Edit .env and train_igpo.sh first
bash train_igpo.sh
See
RL/README.mdfor reward configuration, rollout setup, and troubleshooting.
🔌 External Services
Both Inference/ and RL/ depend on external tools and model endpoints. SFT does not require these.
| Service | Purpose |
|---|---|
| Serper | Web search |
| Jina Reader | Webpage fetching |
| OpenAI-compatible API | Page summarization |
| OpenAI-compatible judge model | RL reward evaluation |
📁 Repository Layout
DR-Venus/
├── README.md
├── assets/
│ ├── dr-venus-overview.png
│ └── dr-venus-passk.png
├── SFT/
│ ├── data_clean/ # RED trajectory conversion & cleaning
│ ├── scripts/ # Checkpoint merge utilities
│ ├── sft_shells/
│ ├── train_sft.sh
│ └── README.md
├── RL/
│ ├── configs/
│ ├── data/
│ ├── tool_server/
│ ├── train_igpo.sh
│ └── README.md
├── Inference/
│ ├── tool_server/
│ ├── run_demo.sh
│ ├── run_web_demo.sh
│ └── README.md
└── model_cards/
🤝 Acknowledgements
DR-Venus builds on several strong open-source foundations:
- verl -- training infrastructure
- IGPO -- RL algorithmic foundation
- Tongyi DeepResearch -- deep research agent design references
- REDSearcher -- open-data trajectories and related tooling
📝 Citation
@article{venus2026drvenus,
title={DR-Venus: Towards Frontier Edge-Scale Deep Research Agents with Only 10K Open Data},
author={Venus Team and Dai, Sunhao and Deng, Yong and Lin, Jinzhen and Song, Yusheng and Wang, Guoqing and Wu, Xiaofeng and Zhou, Yuqi and Yang, Shuo and Ying, Zhenzhe and Zhang, Zhanwei and Meng, Changhua and Wang, Weiqiang},
journal={arXiv preprint arXiv:2604.19859},
year={2026}
}