Mixture-of-Experts

A sparse architecture that routes each token to a small subset of expert sub-networks, so a model with very large total parameters activates only a fraction per forward pass — the key to running frontier-scale models on modest local hardware.

grounded in: 'GLM 5.2 local run' notable item + the dominant trend theme 'Local LLM inference on modest hardware: squeezing frontier models onto slow/small machines' (GLM-class models are MoE)

Connected concepts

Quantization, Local-first inference, Distributed inference, vLLM / PagedAttention