AgentCPM-Report uses a Writing As Reasoning Policy to iteratively refine outlines and draft evidence-grounded long-form research reports.
Highlights
8B Parameter quantity — Extreme performance, extremely small size：Through an average of about 40 rounds of in-depth retrieval and nearly 100 rounds of thought chain reasoning, comprehensive information mining and reconstruction is achieved, allowing the client-side model to produce tens of thousands of words with rigorous logic and profound insights. With only 8 billion parameters, it achieves performance comparable to top closed-source systems on deep research tasks.
Physical isolation, local security：Designed specifically for high-privacy scenarios, it supports completely offline and agile local deployment, completely eliminating the risk of cloud data leakage. With the help of the UltraRAG framework, local private knowledge bases can be efficiently mounted and understood, and core confidential data can be securely transformed into highly valuable professional decision-making reports without leaving the local domain.
Technical highlights
Writing as Reasoning Strategy (WARP)：During the report generation process, the outline is dynamically revised, and evidence-based drafting and reasoning-driven deepening are alternately carried out; the content is iteratively reconstructed through an average of about 40 rounds of in-depth searches and nearly 100 rounds of thinking chain reasoning to generate high-quality, long and logically rigorous research reports.
UltraRAG powered local private knowledge base：Based on the UltraRAG framework, approximately 2.7 million Arxiv papers and local private knowledge bases are mounted, supporting completely offline terminal-side deployment; in-depth retrieval and writing are completed under the premise that the data is physically isolated and does not leave the local domain, and confidential data is securely transformed into professional decision-making reports.
Benchmarks
DeepResearch Bench
DeepResearch Gym
DeepConsult
The evaluation data set covers DeepResearch Bench, DeepConsult and DeepResearch Gym; the knowledge base used for writing contains about 2.7 million Arxiv papers and about 200,000 internal web page abstracts.