Hugging Face Overview
AgentCPM-Explore: A 4B Parameter Agent Foundation Model
AgentCPM-Explore, developed by OpenBMB, is a 4 billion parameter agent foundation model specifically engineered for tackling complex, long-horizon agent benchmarks. It is notable for being the first full-parameter 4B agent model to rank on 8 such benchmarks, including GAIA, HLE, and BrowserComp, demonstrating strong performance in on-device settings.
Key Capabilities
- Deep Exploration: Capable of over 100 rounds of continuous environment interaction, facilitating multi-source information cross-validation and dynamic search strategy adjustments.
- Real-time Verification: Supports real-time verification of up-to-date information, enabling sustained deep exploration until task completion.
- Comprehensive Open-Source Ecosystem: The project includes fully open-sourced components such as AgentRL (an asynchronous reinforcement learning framework), AgentDock (a unified tool sandbox management platform), and AgentToLeaP (a one-click evaluation platform for tool-learning capabilities. These components foster community collaboration and custom extensibility.
Performance Highlights
AgentCPM-Explore 4B demonstrates competitive performance across various agent benchmarks. For instance, it achieves 63.9% on GAIA (text-only), 25.0% on BrowseComp, 82.7% on Frames, and 70.0% on Xbench-DeepSearch, often outperforming or matching larger open-source models in specific tasks while maintaining a significantly smaller parameter count.
Good for
- Developing on-device agents requiring efficient, deep research capabilities.
- Researchers and developers interested in agent foundation models and their training infrastructure.
- Applications demanding long-horizon task completion with continuous environmental interaction and information synthesis.