AgentCPM is an agent model family from OpenBMB, including AgentCPM-Explore for GUI agents and AgentCPM-Report for deep research.
Highlights
4B Parameter scale — End-side deep search agent：The first full-parameter 4 billion-scale agent model successfully ranked on 8 long-term, high-complexity agent benchmarks in the device-side scenario, including GAIA, HLE and BrowserComp.
100+ successive rounds of interaction — Long-term continuous exploration：It can support more than 100 rounds of continuous environment interaction, and has the ability to cross-validate multi-source information, adjust dynamic search strategies, and verify the latest information in real time, thereby achieving continuous and in-depth exploration until the task is completed.
3 Open source core components — Completely open source from end to end：End-to-end fully open source: (1) AgentRL - a fully asynchronous reinforcement learning framework for agent training; (2) AgentDock - a unified tool sandbox management and scheduling platform; (3) AgentToLeaP - a one-click agent tool learning capability evaluation platform. Together, these components support community collaboration and custom extensions.
Technical highlights
Extremely lightweight, superior performance：With only 4B parameters, xbench-DeepSearch 70.0% is the same as the 70B model, and Frames 82.7% is close to the 30B large model, achieving high performance with small and large results.
Outstanding web page perception ability：WebWalkerQA 68.1% surpasses the 8B model and approaches the 30B level. Frames visual understanding is the second open source. In-depth search and web page interaction are core advantage scenarios.
A new benchmark for open source small models：Achieve multi-task availability at the 4B level and provide cost-effective agent solutions for resource-constrained scenarios.
Benchmarks
closed source model
Open source model
The evaluation covers a total of 8 long-term, high-complexity agent benchmarks: GAIA (text only), BrowserComp, BrowseComp (ZH), HLE, Frames, WebWalkerQA, Seal-0, and xbench-DeepSearch; "/" in the table indicates that there is currently no corresponding data for this model.