openbmb/AgentCPM-Report

1---2license: apache-2.03library_name: transformers4pipeline_tag: text-generation5tags:6- agent7- text-generation-inference8---9 10# AgentCPM-Report: Gemini-2.5-pro-DeepResearch Level Local DeepResearch11 12<p align="center">13<a href='https://huggingface.co/openbmb/AgentCPM-Report'><img src='https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-AgentCPM--Report-yellow'>14<a href='https://huggingface.co/openbmb/AgentCPM-Report-GGUF'><img src='https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-AgentCPM--Report--GGUF-yellow'>15<a href='https://github.com/OpenBMB/AgentCPM'><img src='https://img.shields.io/badge/GitHub-AgentCPM-blue?logo=github'>16<a href='https://github.com/OpenBMB/UltraRAG'><img src='https://img.shields.io/badge/GitHub-UltraRAG-blue?logo=github'>17<a href='https://arxiv.org/abs/2602.06540'><img src='https://img.shields.io/badge/arXiv-2602.06540-red'>18</p>19 20This repository contains **AgentCPM-Report**, an 8B-parameter deep research agent introduced in the paper [AgentCPM-Report: Interleaving Drafting and Deepening for Open-Ended Deep Research](https://arxiv.org/abs/2602.06540). 21 22AgentCPM-Report uses a **Writing As Reasoning Policy (WARP)** to dynamically revise outlines during report generation, alternating between evidence-based drafting and reasoning-driven deepening to produce high-quality, long-form research reports.23 24## Links & Resources25### 📊 AgentCPM-Report Models26- **[AgentCPM-Report](https://huggingface.co/openbmb/AgentCPM-Report)** The Gemini-2.5-pro-DeepResearch Level Local DeepResearch Model27- **[AgentCPM-Report-GGUF](https://huggingface.co/openbmb/AgentCPM-Report-GGUF)** The GGUF version of AgentCPM-Report28 29### 🤖 AgentCPM-Explore Models  30- **[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.31- **[AgentCPM-Explore-GGUF](https://huggingface.co/openbmb/AgentCPM-Explore-GGUF)** The GGUF version of AgentCPM-Explore32  33### 💻 Code & Framework34- **[AgentCPM](https://github.com/OpenBMB/AgentCPM)** Our code for AgentCPM Series35- **[UltraRAG](https://github.com/OpenBMB/UltraRAG)** A RAG Framework, Less Code, Lower Barrier, Faster Deployment36 37 38## News39- [2026-01-20] 🚀🚀🚀 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.40 41## Overview42AgentCPM-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:43 44- **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.  45- **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.46 47## Demo Cases48<div align="center">49  <a href="https://www.youtube.com/watch?v=d5XWONt0PWo"><img src="https://img.youtube.com/vi/d5XWONt0PWo/0.jpg", width=70%></a>50</div>51 52**You can watch our demo video here [Demo](https://www.youtube.com/watch?v=d5XWONt0PWo) 🔗**53 54 55## Quick Start56### Docker Deployment57<div align="center">58  <a href="https://www.youtube.com/watch?v=ze8qJRrass4"><img src="https://img.youtube.com/vi/ze8qJRrass4/0.jpg", width=70%></a>59</div>60 61**You can watch our demo video here [Tutorial](https://www.youtube.com/watch?v=ze8qJRrass4) 🔗**62 63We 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 models—just switch `docker-compose.yml` to `docker-compose.cpu.yml`.64 65``` bash66git clone git@github.com:OpenBMB/UltraRAG.git67cd UltraRAG68git checkout agentcpm-report-demo69cd agentcpm-report-demo70cp env.example .env71docker-compose -f docker-compose.yml up -d --build72docker-compose -f docker-compose.yml logs -f ultrarag-ui73``` 74The first startup pulls images, downloads the model, and configures the environment, which takes about 30 minutes.75Then open `http://localhost:5050`. If you can see the UI, your deployment is successful.76Follow 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.77 78(Optional) You can import [Wiki2024](https://modelscope.cn/datasets/UltraRAG/UltraRAG_Benchmark/tree/master/corpus/wiki24) as the writing database.79 80You can read more tutorials about AgentCPM-Report in the [documentation](https://ultrarag.openbmb.cn/pages/en/demo/deepresearch).81 82 83## Evaluation84<table align="center">85  <thead>86    <tr>87      <th align="center">DeepResearch Bench</th>88      <th align="center">Overall</th>89      <th align="center">Comprehensiveness</th>90      <th align="center">Insight</th>91      <th align="center">Instruction Following</th>92      <th align="center">Readability</th>93    </tr>94  </thead>95  <tbody>96    <tr>97      <td align="center">Doubao-research</td>98      <td align="center">44.34</td>99      <td align="center">44.84</td>100      <td align="center">40.56</td>101      <td align="center">47.95</td>102      <td align="center">44.69</td>103    </tr>104    <tr>105      <td align="center">Claude-research</td>106      <td align="center">45.00</td>107      <td align="center">45.34</td>108      <td align="center">42.79</td>109      <td align="center">47.58</td>110      <td align="center">44.66</td>111    </tr>112    <tr>113      <td align="center">OpenAI-deepresearch</td>114      <td align="center">46.45</td>115      <td align="center">46.46</td>116      <td align="center">43.73</td>117      <td align="center">49.39</td>118      <td align="center">47.22</td>119    </tr>120    <tr>121      <td align="center">Gemini-2.5-Pro-deepresearch</td>122      <td align="center">49.71</td>123      <td align="center">49.51</td>124      <td align="center">49.45</td>125      <td align="center">50.12</td>126      <td align="center">50.00</td>127    </tr>128    <tr>129      <td align="center">WebWeaver(Qwen3-30B-A3B)</td>130      <td align="center">46.77</td>131      <td align="center">45.15</td>132      <td align="center">45.78</td>133      <td align="center">49.21</td>134      <td align="center">47.34</td>135    </tr>136    <tr>137      <td align="center">WebWeaver(Claude-Sonnet-4)</td>138      <td align="center">50.58</td>139      <td align="center">51.45</td>140      <td align="center">50.02</td>141      <td align="center">50.81</td>142      <td align="center">49.79</td>143    </tr>144    <tr>145      <td align="center">Enterprise-DR(Gemini-2.5-Pro)</td>146      <td align="center">49.86</td>147      <td align="center">49.01</td>148      <td align="center">50.28</td>149      <td align="center">50.03</td>150      <td align="center">49.98</td>151    </tr>152    <tr>153      <td align="center">RhinoInsigh(Gemini-2.5-Pro)</td>154      <td align="center">50.92</td>155      <td align="center">50.51</td>156      <td align="center">51.45</td>157      <td align="center">51.72</td>158      <td align="center">50.00</td>159    </tr>160    <tr>161      <td align="center">AgentCPM-Report</td>162      <td align="center">50.11</td>163      <td align="center">50.54</td>164      <td align="center">52.64</td>165      <td align="center">48.87</td>166      <td align="center">44.17</td>167    </tr>168  </tbody>169</table>170 171 172 173<table align="center">174  <thead>175    <tr>176      <th align="center">DeepResearch Gym</th>177      <th align="center">Avg.</th>178      <th align="center">Clarity</th>179      <th align="center">Depth</th>180      <th align="center">Balance</th>181      <th align="center">Breadth</th>182      <th align="center">Support</th>183      <th align="center">Insightfulness</th>184    </tr>185  </thead>186  <tbody>187    <tr>188      <td align="center">Doubao-research</td>189      <td align="center">84.46</td>190      <td align="center">68.85</td>191      <td align="center">93.12</td>192      <td align="center">83.96</td>193      <td align="center">93.33</td>194      <td align="center">84.38</td>195      <td align="center">83.12</td>196    </tr>197    <tr>198      <td align="center">Claude-research</td>199      <td align="center">80.25</td>200      <td align="center">86.67</td>201      <td align="center">96.88</td>202      <td align="center">84.41</td>203      <td align="center">96.56</td>204      <td align="center">26.77</td>205      <td align="center">90.22</td>206    </tr>207    <tr>208      <td align="center">OpenAI-deepresearch</td>209      <td align="center">91.27</td>210      <td align="center">84.90</td>211      <td align="center">98.10</td>212      <td align="center">89.80</td>213      <td align="center">97.40</td>214      <td align="center">88.40</td>215      <td align="center">89.00</td>216    </tr>217    <tr>218      <td align="center">Gemini-2.5-pro-deepresearch</td>219      <td align="center">96.02</td>220      <td align="center">90.71</td>221      <td align="center">99.90</td>222      <td align="center">93.37</td>223      <td align="center">99.69</td>224      <td align="center">95.00</td>225      <td align="center">97.45</td>226    </tr>227    <tr>228      <td align="center">WebWeaver (Qwen3-30b-a3b)</td>229      <td align="center">77.27</td>230      <td align="center">71.88</td>231      <td align="center">85.51</td>232      <td align="center">75.80</td>233      <td align="center">84.78</td>234      <td align="center">63.77</td>235      <td align="center">81.88</td>236    </tr>237    <tr>238      <td align="center">WebWeaver (Claude-sonnet-4)</td>239      <td align="center">96.77</td>240      <td align="center">90.50</td>241      <td align="center">99.87</td>242      <td align="center">94.30</td>243      <td align="center">100.00</td>244      <td align="center">98.73</td>245      <td align="center">97.22</td>246    </tr>247    <tr>248      <td align="center">AgentCPM-Report</td>249      <td align="center">98.48</td>250      <td align="center">95.10</td>251      <td align="center">100.00</td>252      <td align="center">98.50</td>253      <td align="center">100.00</td>254      <td align="center">97.30</td>255      <td align="center">100.00</td>256    </tr>257  </tbody>258</table>259 260<table align="center">261  <thead>262    <tr>263      <th align="center">DeepConsult</th>264      <th align="center">Avg.</th>265      <th align="center">Win</th>266      <th align="center">Tie</th>267      <th align="center">Lose</th>268    </tr>269  </thead>270  <tbody>271    <tr>272      <td align="center">Doubao-research</td>273      <td align="center">5.42</td>274      <td align="center">29.95</td>275      <td align="center">40.35</td>276      <td align="center">29.70</td>277    </tr>278    <tr>279      <td align="center">Claude-research</td>280      <td align="center">4.60</td>281      <td align="center">25.00</td>282      <td align="center">38.89</td>283      <td align="center">36.11</td>284    </tr>285    <tr>286      <td align="center">OpenAI-deepresearch</td>287      <td align="center">5.00</td>288      <td align="center">0.00</td>289      <td align="center">100.00</td>290      <td align="center">0.00</td>291    </tr>292    <tr>293      <td align="center">Gemini-2.5-Pro-deepresearch</td>294      <td align="center">6.70</td>295      <td align="center">61.27</td>296      <td align="center">31.13</td>297      <td align="center">7.60</td>298    </tr>299    <tr>300      <td align="center">WebWeaver(Qwen3-30B-A3B)</td>301      <td align="center">4.57</td>302      <td align="center">28.65</td>303      <td align="center">34.90</td>304      <td align="center">36.46</td>305    </tr>306    <tr>307      <td align="center">WebWeaver(Claude-Sonnet-4)</td>308      <td align="center">6.96</td>309      <td align="center">66.86</td>310      <td align="center">10.47</td>311      <td align="center">22.67</td>312    </tr>313    <tr>314      <td align="center">Enterprise-DR(Gemini-2.5-Pro)</td>315      <td align="center">6.82</td>316      <td align="center">71.57</td>317      <td align="center">19.12</td>318      <td align="center">9.31</td>319    </tr>320    <tr>321      <td align="center">RhinoInsigh(Gemini-2.5-Pro)</td>322      <td align="center">6.82</td>323      <td align="center">68.51</td>324      <td align="center">11.02</td>325      <td align="center">20.47</td>326    </tr>327    <tr>328      <td align="center">AgentCPM-Report</td>329      <td align="center">6.60</td>330      <td align="center">57.60</td>331      <td align="center">13.73</td>332      <td align="center">28.68</td>333    </tr>334  </tbody>335</table>336 337Our 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.338 339## Acknowledgements340This 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/).341 342## Contributions343Project leads: Yishan Li, Wentong Chen344 345Contributors: 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 Sun346 347Advisors: Yukun Yan, Yankai Lin, Zhiyuan Liu, Maosong Sun348 349## Citation350 351If **AgentCPM-Report** is helpful for your research, please cite it as follows:352 353```bibtex354@misc{li2026agentcpmreport,355      title={AgentCPM-Report: Interleaving Drafting and Deepening for Open-Ended Deep Research}, 356      author={Yishan Li and Wentong Chen and Yukun Yan and Mingwei Li and Sen Mei and Xiaorong Wang and Kunpeng Liu and Xin Cong and Shuo Wang and Zhong Zhang and Yaxi Lu and Zhenghao Liu and Yankai Lin and Zhiyuan Liu and Maosong Sun},357      year={2026},358      eprint={2602.06540},359      archivePrefix={arXiv},360      primaryClass={cs.AI},361      url={https://arxiv.org/abs/2602.06540}, 362}363```