Researchers propose AI-ModelNet: an Internet for AI models

In a new paper on arXiv, Li Zhetao and colleagues introduce AI-ModelNet, a concept that treats AI models as nodes in a global network, akin to the internet. The researchers argue that while large models (LMs) excel at complex tasks, their high training costs and deployment hurdles push the field toward smaller, private, domain-specific models. The challenge then becomes enabling these heterogeneous models to interact and collaborate effectively. AI-ModelNet proposes a hierarchical architecture with standard protocols for model discovery, communication, and collaborative reasoning, allowing models to request capabilities from each other on demand.

The paper reviews current single-model and multi-model research, then details a prototype system that demonstrates feasibility across diverse applications such as cross-model knowledge transfer and distributed inference. Key future directions include model routing, security, and incentive mechanisms. By repurposing the Internet's principles of openness and interconnection, AI-ModelNet could dramatically lower the barrier to leveraging AI, enabling specialized models to pool their strengths without requiring a single massive infrastructure investment.