DeepResearch finally goes local: AgentCPM-Report, an 8B on-device writing agent, is now open source
With deep research sweeping the industry, we all want a personal “super writing assistant” that can synthesize complex information and draft ten-thousand-word reports on its own. But when you are holding next year’s corporate strategy, unreleased financial statements, or research data involving core secrets, would you really dare upload them to the cloud?
This is the biggest problem in deep research today: top-tier deep research capabilities require cloud large models, which means exposing important data; absolute data security means going offline or using small local models, whose reports are often shallow and of little practical use.
Must data privacy and deep insight be mutually exclusive?
Today we offer a way out: AgentCPM-Report, a deep research agent that is local, private, and still SOTA. It was jointly developed by the Tsinghua University NLP lab, Renmin University of China, ModelBest, and the OpenBMB open-source community. Built around an on-device model, it delivers report-writing capability on par with top closed-source systems. That means no expensive compute cluster and no uploading information over the network — you get an expert-level research assistant right on your own machine.
💡 AgentCPM-Report highlights
Extreme efficiency, punching far above its weight: through an average of 40 rounds of deep retrieval and nearly 100 rounds of chain-of-thought reasoning, it mines and restructures information from every angle, letting on-device models produce tightly argued, deeply insightful long-form reports and matching top closed-source systems on deep research tasks at 8B parameters.
Physically isolated and locally secure: designed for high-privacy scenarios, it supports fully offline, agile local deployment that eliminates any risk of cloud leaks. Built on our UltraRAG framework, it efficiently mounts and understands your local private knowledge base, so core confidential data can be turned into valuable professional decision reports without ever leaving your domain.
➤ UltraRAG framework repository
🔗 https://github.com/OpenBMB/UltraRAG
➤ AgentCPM-Report repositories
GitHub:
🔗 https://github.com/OpenBMB/AgentCPM
HuggingFace:
🔗 https://huggingface.co/openbmb/AgentCPM-Report
ModelScope:
🔗https://modelscope.cn/models/OpenBMB/AgentCPM-Report
GitCode:
🔗 https://gitcode.com/OpenBMB/AgentCPM
Modelers:
🔗 https://modelers.cn/models/OpenBMB/AgentCPM-Report
Writing quality measured against the strongest closed-source benchmarks: on the three mainstream deep research benchmarks — DeepResearch Bench, Deep Consult, and DeepResearch Gym — AgentCPM-Report shows remarkable cross-tier strength, with overall scores matching or exceeding top closed-source systems. On the insight metric, which best tests depth, AgentCPM-Report ranks first ahead of all competitors; on comprehensiveness, it trails only a Claude-based complex writing framework and sits in the top tier.
Building a deep report on the “Wallfacer Project” from The Three-Body Problem from scratch
Benchmark scores are not enough, so we put it to work. We asked AgentCPM-Report to use the original text of The Three-Body Problem as its knowledge base and produce an investigative report on the origins and evolution of the “Wallfacer Project.” In the video below, you will see the entire process, from digging up clues and planning an outline to writing a ten-thousand-word report:
This article includes video content. Please view it in the original WeChat post.
Minimal deployment guide: one-click deployment with the UltraRAG framework gives you your own deep research assistant:
One-click Docker startup: launch the UltraRAG service and the AgentCPM agent locally in seconds with Docker;
Drag-and-drop knowledge base construction: no code required. Drag local PDFs, TXT files, and other private documents into the backend, and the system handles chunking and vector indexing automatically;
Immersive deep research: enter a research topic and let the agent generate a structured, cited professional report.
Technical deep dive: how an on-device model wins against bigger systems
AgentCPM-Report can match closed-source systems with 8B parameters thanks to two technical innovations.
Innovation 1: a “writing as reasoning” paradigm that thinks more like a human brain. Conventional approaches try to make the model generate a complete outline or body of content in one pass, which easily causes the report’s logic to collapse — and is even harder for small models. AgentCPM-Report breaks this limitation with an iterative refinement framework that plans while writing:
Two-stage loop: the system alternates between “drafting” and “deepening.” Like a human expert, after finishing a draft section it pauses to reflect: “Is anything missing? What should the focus be?” Then it goes back to expand sections, run additional retrieval, and fill in new content.
Progressive optimization: the massive long-form writing task is broken into a series of small, executable goals. Each loop only has to solve the local problem at hand, allowing a smaller model to produce long reports that are logically tight and rich in detail.
Innovation 2: multi-stage agent learning for all-around capability gains: 01 Breaking down four core capabilities
We decomposed full report collaboration into four core capability modules and trained each one specifically:
🔍 Intelligent retrieval: recall as the primary optimization metric, ensuring highly relevant retrieved content and a solid foundation for writing;
📝 Fluent writing: a multi-dimensional quality evaluation system that checks everything from depth of content to clarity of expression;
🗺️ Sound planning: structural evaluation of generated outlines to ensure rigorous logic and clear hierarchy;
🎯 Precise decision-making: trajectory pruning to construct high-quality data, solving the key decision of when to stop deepening.
02 A three-stage training recipe: from beginner to expert
Our training process is carefully designed as well:
Supervised fine-tuning: guided by high-quality exemplars to learn the basic writing paradigm;
Atomic capability reinforcement: targeted improvement for each core atomic capability;
Full-pipeline optimization: end-to-end, full-chain reinforcement learning with final report quality as the sole objective, connecting all capabilities.
Try it now: put DeepResearch on your own hard drive
AgentCPM-Report is now open source. It is the first deep research system that is truly deployable locally and practical enough for real use. Your data belongs only to you. We invite developers and researchers to follow our tutorial, deploy it, load your private knowledge base, and experience professional-grade report generation where data never leaves your domain!
📧 Partnership contact: agentcpm@modelbest.cn;
💬 Technical discussion: scan the QR code below to join the group. We look forward to your suggestions and feedback.