How a 4B‑Parameter AgentCPM‑Explore Beats 30B Models in Long‑Range Tasks

Core Highlights: Redefining Small Model Capabilities

1. Breaking Parameter Constraints, Reshaping Performance Ceiling

As the first 4B‑parameter model that can handle eight long‑term intelligent‑agent tasks such as GAIA, Xbench, Browsercomp, AgentCPM‑Explore disproves the notion that larger parameters guarantee stronger performance. Within a limited parameter budget it achieves a leap in capability, opening new pathways for small‑model development.

2. Unlimited Long‑Range Exploration, Stable Interaction

When faced with complex, long‑duration tasks, the model demonstrates strong sustained exploration, delivering over 100 rounds of non‑repeating, stable environment interactions. It maintains coherent reasoning even for intricate tasks, addressing the common interruption and error issues of small models in long‑term interaction.

3. Full‑Process Open‑Source Enablement, Flexible Extension

Unlike many projects that only open‑source the model itself, AgentCPM‑Explore provides a complete ecosystem: the AgentDock sandbox for unified tool management, the asynchronous reinforcement‑learning framework AgentRL, and the one‑click evaluation platform AgentToLeaP. This allows developers to reproduce training end‑to‑end and customize extensions, lowering the R&D barrier for small‑model agents.

Performance Benchmarks: Small Model, Big Gains

Across eight mainstream agent benchmarks—including GAIA, HLE, Browsercomp, WebWalker—AgentCPM‑Explore shows exceptional parameter‑efficiency. On the Xbench‑DeepResearch task it surpasses closed‑source models such as Claude‑4.5‑Sonnet and OpenAI‑o3, exceeding the 8B‑model trend line by 9.4% and the 4B‑model trend line by 35.5%.

Specifically, on the GAIA text subset the model scores 63.9%, far above same‑size models; on Frames it achieves 82.7%, ranking in the top tier just behind Tongyi deep‑research 30B; on WebWalker it reaches 68.1%, beating Mirothinker 8B. The evaluation uses a high‑standard Avg@8 metric, keeping sampling variance under 2%.

With the support of AgentDock and AgentRL, the model solves over 95% of GAIA text tasks when multiple attempts are allowed, overturning the stereotype that small models are inherently limited.

Intelligent Behavior: Near‑Human Reasoning

In deep‑exploration tasks, AgentCPM‑Explore exhibits reasoning akin to human thought, as illustrated by its handling of a complex question about the farthest‑apart US presidential birth cities.

• Questioning: It does not blindly accept tool output; when “Brookline, MA” is flagged as the easternmost city, it suspects missing information and requests a full data check.

• Pursuing Truth: It refuses compressed secondary data, seeking the original full dataset to base decisions on complete facts.

• Adapting: When generic search fails, it switches to table‑scraping and, if paths mismatch, locates precise GitHub resources to refine the plan.

• Persistence: It continues probing alternative sources despite repeated dead‑ends until reliable data is found.

Open‑Source Infrastructure: Three Core Tools for Efficient Development

The success of AgentCPM‑Explore relies on three open‑source building blocks that together form an efficient R&D system for small‑model agents.

1. AgentDock: Unified Tool‑Sandbox Management Platform

The platform natively supports 16 MCP services and hundreds of tools, offering multi‑version polling and load‑balancing to achieve over 100 QPS for core high‑frequency tools. It includes fault‑tolerance, output standardization, automatic retries, self‑healing, and hot‑swap tool capabilities, handling task distribution, container orchestration, and dynamic routing so developers can focus on capability interfaces.

2. AgentRL: Minimalist Asynchronous Reinforcement‑Learning Framework

AgentRL’s core code spans only seven files (~1,000 lines), lowering the learning curve. It integrates via a standard ChatCompletions API, runs training and sampling asynchronously on the same GPU, and decouples sampling from training, supporting PyTorch native parallelism and advanced techniques such as FSDP2, Tensor Parallel, and Context Parallel for 128K+ token training.

3. AgentToLeaP: One‑Click Agent Capability Evaluation Platform

AgentToLeaP enables one‑click evaluation on eight major leaderboards (GAIA, HLE, etc.). With a single command it launches full‑process testing, offering modular test‑set management and unified result output, allowing developers to plug in custom datasets for precise model optimization.

Technical Insights: How a Small Model “Wins Big”

With only 4 billion parameters, the model faces limited fault tolerance in long‑term, multi‑interaction tasks. The research team identified three key challenges and proposed concrete solutions.

1. Model Fusion: Overcoming SFT Over‑fitting

During supervised fine‑tuning, small models tend to memorize task‑specific prompts, degrading general decision ability. The team applied parameter fusion, blending the specialized post‑training model with the pre‑training generic model. This preserves shared generalization parameters while enhancing specialized skills, yielding about a 7% performance boost on agent tasks.

2. Signal Denoising: Correcting RL Reward Bias

Long trajectories contain many steps, and negative signals at the final step can corrupt earlier correct reasoning. By filtering reward signals and avoiding full‑trajectory penalties for failed end states, the team prevents harmful noise from damaging the model’s learned logic.

3. Information Refinement: Counteracting Long‑Text Interference

Web‑derived noisy information can derail reasoning. The team introduced a context‑refinement mechanism that uses auxiliary models or multi‑model collaboration to summarize and filter web content before feeding it to the 4B model, improving GAIA performance by up to 10%.

AgentCPM‑Explore demonstrates the huge potential of small models in parameter efficiency and, through a fully open‑source ecosystem, paves the way for edge‑side intelligent agents. Continued iteration promises broader deployment of compact AI capabilities.

Open‑source repository: https://github.com/OpenBMB/AgentCPM