How a Local 8B LLM Beats Closed‑Source Giants in Deep Research

Problem Statement

Deep research tasks require synthesizing large amounts of confidential information into long, high‑quality reports. Cloud‑based large language models either expose sensitive data or produce shallow, unreliable outputs, while small offline models lack the capability to generate insight‑rich documents.

AgentCPM-Report Overview

AgentCPM-Report is a fully local, privacy‑preserving research agent built on an 8‑billion‑parameter end‑side model. It achieves state‑of‑the‑art (SOTA) report‑writing performance comparable to leading closed‑source systems without any network communication.

Key Highlights

Extreme Efficiency : An average of 40 retrieval rounds and ~100 reasoning steps enable the model to produce logically rigorous, multi‑thousand‑word reports, matching top‑tier closed‑source quality with only 8B parameters.

Physical Isolation & Security : Designed for high‑privacy scenarios, the system runs completely offline via the UltraRAG framework, ingesting private knowledge bases without any data leaving the local environment.

Technical Innovations

Writing as Reasoning Loop : Instead of generating a full outline in one pass, the system iteratively drafts and refines sections, alternating between a “draft” state and a “deepening” state. This two‑stage cycle mirrors expert workflows, allowing the model to reflect, expand, and improve each segment before proceeding.

Multi‑Stage Agent Learning : The model is trained on four core abilities—intelligent retrieval, fluent writing, scientific planning, and precise decision‑making—through a three‑phase curriculum: supervised fine‑tuning with high‑quality exemplars, atomic ability reinforcement, and end‑to‑end full‑pipeline reinforcement learning where final report quality is the sole objective.

Training Pipeline

Supervised fine‑tuning on curated reports to teach basic writing patterns.

Atomic ability strengthening for each of the four core capabilities.

Full‑process reinforcement learning that links all ability modules and optimizes the final report score.

Benchmark Performance

On the DeepResearch Bench, Deep Consult, and DeepResearch Gym benchmarks, AgentCPM-Report attains a composite score that surpasses most closed‑source systems. It ranks first on the most demanding insight metrics and remains in the top tier for comprehensiveness, comparable to Claude‑based frameworks.

Deployment

Docker One‑Click Start : A provided Docker image launches UltraRAG services and the AgentCPM agent locally within seconds.

Drag‑and‑Drop Knowledge Base : Private PDFs, TXT files, etc., can be imported without code; the system automatically chunks and vectorizes them.

Immersive Deep Research : Users input a research topic and receive a structured, citation‑rich professional report.

Open‑Source Resources

UltraRAG framework: https://github.com/OpenBMB/UltraRAG

AgentCPM‑Report code: https://github.com/OpenBMB/AgentCPM

Model on HuggingFace: https://huggingface.co/openbmb/AgentCPM-Report

ModelScope entry: https://modelscope.cn/models/OpenBMB/AgentCPM-Report

Example Usage

When provided with the novel Three‑Body Problem as a knowledge base, AgentCPM-Report can autonomously generate a detailed investigation report on the “Wall‑Facing Plan” storyline, demonstrating end‑to‑end capabilities from clue extraction and outline planning to full‑length writing.