---
license: apache-2.0
---
# AgentCPM-Report: Gemini-2.5-pro-DeepResearch Level Local DeepResearch
## Links & Resources
### π AgentCPM-Report Models
- **[AgentCPM-Report](https://huggingface.co/openbmb/AgentCPM-Report)** The Gemini-2.5-pro-DeepResearch Level Local DeepResearch Model
- **[AgentCPM-Report-GGUF](https://huggingface.co/openbmb/AgentCPM-Report-GGUF)** The GGUF version of AgentCPM-Report
### π€ AgentCPM-Explore Models
- **[AgentCPM-Explore](https://huggingface.co/openbmb/AgentCPM-Explore)** The first open-source agent model with 4B parameters to appear on 8 widely used long-horizon agent benchmarks.
- **[AgentCPM-Explore-GGUF](https://huggingface.co/openbmb/AgentCPM-Explore-GGUF)** The GGUF version of AgentCPM-Explore
### π» Code & Framework
- **[AgentCPM](https://github.com/OpenBMB/AgentCPM)** Our code for AgentCPM Series
- **[UltraRAG](https://github.com/OpenBMB/UltraRAG)** A RAG Framework, Less Code, Lower Barrier, Faster Deployment
## News
- [2026-01-20] πππ We open-sourced AgentCPM-Report built on MiniCPM4.1-8B, capable of matching top closed-source commercial systems like Gemini-2.5-pro-DeepResearch in report generation.
## Overview
AgentCPM-Report is an open-source large language model agent jointly developed by [THUNLP](https://nlp.csai.tsinghua.edu.cn), Renmin University of China [RUCBM](https://github.com/RUCBM), and [ModelBest](https://modelbest.cn/en). It is based on the [MiniCPM4.1](https://github.com/OpenBMB/MiniCPM) 8B-parameter base model. It accepts user instructions as input and autonomously generates long-form reports. Key highlights:
- **Extreme Performance, Minimal Footprint**: Through an average of 40 rounds of deep retrieval and nearly 100 rounds of chain-of-thought reasoning, it achieves comprehensive information mining and restructuring, enabling edge-side models to produce logically rigorous, deeply insightful long-form articles with tens of thousands of words. With just 8 billion parameters, it delivers performance on par with top-tier closed-source systems in deep research tasks.
- **Physical Isolation, Local Security**: Specifically designed for high-privacy scenarios, it supports fully offline and agile local deployment, completely eliminating the risk of cloud data leaks. Leveraging our UltraRAG framework, it efficiently mounts and understands your local private knowledge base, securely transforming core confidential data into highly valuable professional decision-making reports without ever leaving its domain.
## Demo Cases
**You can watch our demo video here [Demo](https://www.youtube.com/watch?v=d5XWONt0PWo) π**
## Quick Start
### Docker Deployment
**You can watch our demo video here [Tutorial](https://www.youtube.com/watch?v=ze8qJRrass4) π**
We provide a minimal one-click `docker-compose` deployment integrated with UltraRAG, including the RAG framework UltraRAG2.0, the model inference framework vllm, and the vector database milvus. If you want CPU inference, we also provide a llama.cpp-based version for gguf modelsβjust switch `docker-compose.yml` to `docker-compose.cpu.yml`.
``` bash
git clone git@github.com:OpenBMB/UltraRAG.git
cd UltraRAG
git checkout agentcpm-report-demo
cd agentcpm-report-demo
cp env.example .env
docker-compose -f docker-compose.yml up -d --build
docker-compose -f docker-compose.yml logs -f ultrarag-ui
```
The first startup pulls images, downloads the model, and configures the environment, which takes about 30 minutes.
Then open `http://localhost:5050`. If you can see the UI, your deployment is successful.
Follow the UI instructions to upload local files, chunk them, and build indexes; then in the Chat section, select AgentCPM-Report in the pipeline to start your workflow.
(Optional) You can import [Wiki2024](https://modelscope.cn/datasets/UltraRAG/UltraRAG_Benchmark/tree/master/corpus/wiki24) as the writing database.
You can read more tutorials about AgentCPM-Report in the [documentation](https://ultrarag.openbmb.cn/pages/en/demo/deepresearch).
## Evaluation
| DeepResearch Bench |
Overall |
Comprehensiveness |
Insight |
Instruction Following |
Readability |
| Doubao-research |
44.34 |
44.84 |
40.56 |
47.95 |
44.69 |
| Claude-research |
45.00 |
45.34 |
42.79 |
47.58 |
44.66 |
| OpenAI-deepresearch |
46.45 |
46.46 |
43.73 |
49.39 |
47.22 |
| Gemini-2.5-Pro-deepresearch |
49.71 |
49.51 |
49.45 |
50.12 |
50.00 |
| WebWeaver(Qwen3-30B-A3B) |
46.77 |
45.15 |
45.78 |
49.21 |
47.34 |
| WebWeaver(Claude-Sonnet-4) |
50.58 |
51.45 |
50.02 |
50.81 |
49.79 |
| Enterprise-DR(Gemini-2.5-Pro) |
49.86 |
49.01 |
50.28 |
50.03 |
49.98 |
| RhinoInsigh(Gemini-2.5-Pro) |
50.92 |
50.51 |
51.45 |
51.72 |
50.00 |
| AgentCPM-Report |
50.11 |
50.54 |
52.64 |
48.87 |
44.17 |
| DeepResearch Gym |
Avg. |
Clarity |
Depth |
Balance |
Breadth |
Support |
Insightfulness |
| Doubao-research |
84.46 |
68.85 |
93.12 |
83.96 |
93.33 |
84.38 |
83.12 |
| Claude-research |
80.25 |
86.67 |
96.88 |
84.41 |
96.56 |
26.77 |
90.22 |
| OpenAI-deepresearch |
91.27 |
84.90 |
98.10 |
89.80 |
97.40 |
88.40 |
89.00 |
| Gemini-2.5-pro-deepresearch |
96.02 |
90.71 |
99.90 |
93.37 |
99.69 |
95.00 |
97.45 |
| WebWeaver (Qwen3-30b-a3b) |
77.27 |
71.88 |
85.51 |
75.80 |
84.78 |
63.77 |
81.88 |
| WebWeaver (Claude-sonnet-4) |
96.77 |
90.50 |
99.87 |
94.30 |
100.00 |
98.73 |
97.22 |
| AgentCPM-Report |
98.48 |
95.10 |
100.00 |
98.50 |
100.00 |
97.30 |
100.00 |
| DeepConsult |
Avg. |
Win |
Tie |
Lose |
| Doubao-research |
5.42 |
29.95 |
40.35 |
29.70 |
| Claude-research |
4.60 |
25.00 |
38.89 |
36.11 |
| OpenAI-deepresearch |
5.00 |
0.00 |
100.00 |
0.00 |
| Gemini-2.5-Pro-deepresearch |
6.70 |
61.27 |
31.13 |
7.60 |
| WebWeaver(Qwen3-30B-A3B) |
4.57 |
28.65 |
34.90 |
36.46 |
| WebWeaver(Claude-Sonnet-4) |
6.96 |
66.86 |
10.47 |
22.67 |
| Enterprise-DR(Gemini-2.5-Pro) |
6.82 |
71.57 |
19.12 |
9.31 |
| RhinoInsigh(Gemini-2.5-Pro) |
6.82 |
68.51 |
11.02 |
20.47 |
| AgentCPM-Report |
6.60 |
57.60 |
13.73 |
28.68 |
Our evaluation datasets include DeepResearch Bench, DeepConsult, and DeepResearch Gym. The writing-time knowledge base includes about 2.7 million [Arxiv papers](https://www.kaggle.com/api/v1/datasets/download/Cornell-University/arxiv) and about 200,000 internal webpage summaries.
## Acknowledgements
This project would not be possible without the support and contributions of the open-source community. During development, we referred to and used multiple excellent open-source frameworks, models, and data resources, including [verl](https://github.com/volcengine/verl), [UltraRAG](https://github.com/OpenBMB/UltraRAG), [MiniCPM4.1](https://github.com/OpenBMB/MiniCPM), and [SurveyGo](https://surveygo.modelbest.cn/).
## Contributions
Project leads: Yishan Li, Wentong Chen
Contributors: Yishan Li, Wentong Chen, Yukun Yan, Mingwei Li, Sen Mei, Xiaorong Wang, Kunpeng Liu, Xin Cong, Shuo Wang, Zhong Zhang, Yaxi Lu, Zhenghao Liu, Yankai Lin, Zhiyuan Liu, Maosong Sun
Advisors: Yukun Yan, Yankai Lin, Zhiyuan Liu, Maosong Sun
## Citation
If **AgentCPM-Report** is helpful for your research, please cite it as follows:
```bibtex
@software{AgentCPMReport2026,
title = {AgentCPM-Report: Gemini-2.5-pro-DeepResearch Level Local DeepResearch},
author = {Yishan Li, Wentong Chen, Yukun Yan, Mingwei Li, Sen Mei, Xiaorong Wang, Kunpeng Liu, Xin Cong, Shuo Wang, Zhong Zhang, Yaxi Lu, Zhenghao Liu, Yankai Lin, Zhiyuan Liu, Maosong Sun},
year = {2026},
url = {https://github.com/OpenBMB/AgentCPM}
}
```