4B AgentCPM-Explore reaches Hugging Face TOP2!
Congratulations to AgentCPM-Explore: less than a week after release, it has quickly caught the attention of our open-source community partners and climbed into the Hugging Face TOP2 🚀!
🤔 Why 4B parameters?
While the industry’s attention is still on larger parameter counts, we chose to explore a different path: how to achieve the advanced capabilities of a deep research agent within a very small parameter budget (4B).
Core capabilities at a glance 📖
✅ Deep exploration and planning: supports more than 100 rounds of stable environment interaction, autonomously completing complex information retrieval and reasoning tasks.
✅ Beyond scale limits: on eight mainstream agent benchmarks including GAIA and Xbench, it outperforms some 30B+ models and closed-source models (such as Claude-4.5-sonnet).
✅ Full-pipeline open source: we open-source not only the model weights but also the training framework AgentRL, the tool platform AgentDock, and the evaluation toolchain, supporting full reproduction and secondary development.
🙌 Key breakthrough: the real bottleneck in system performance is the information flow
We are the first to propose and validate the “information bottleneck” hypothesis: in a multi-module agent system, the performance ceiling is set by the narrowest link in the information chain. By improving information throughput and compression fidelity, we broke through previous performance ceilings and offered a new theoretical direction for on-device agents.
Open source, for better co-creation 🤝!
We believe the future of agents belongs to openness and collaboration. We invite every developer and researcher to join us in redefining what on-device agents can do ✨
➤ Explore now
GitHub:
🔗 https://github.com/OpenBMB/AgentCPM
HuggingFace:
🔗 https://huggingface.co/openbmb/AgentCPM-Explore
ModelScope:
🔗 https://modelscope.cn/models/OpenBMB/AgentCPM-Explore
GitCode:
🔗 https://gitcode.com/OpenBMB/AgentCPM
Modelers:
🔗 https://modelers.cn/models/OpenBMB/AgentCPM-Explore