Expert Selection System for MoE Large Language Models

USPTO published patent application US20260099697A1 disclosing a system and method for expert selection in Mixture of Experts (MoE) large language models. The invention comprises a global selector configured to operate in either global mode (selecting a global expert set for each layer prior to inference) or local mode, along with a pre-fetcher that moves selected global expert sets from main memory into faster cache memory. The system identifies and pre-loads "global hot experts" expected to be frequently accessed.

Affected parties include AI researchers, LLM developers, semiconductor manufacturers designing AI accelerators, and cloud computing providers deploying large language models. The patent addresses computational efficiency challenges in MoE architectures where multiple specialized expert networks must be selectively activated. Organizations developing or deploying MoE-based AI systems may wish to monitor this application for potential licensing considerations or design-around opportunities.