Concept 11 of 19
frameworkMixture of Experts (MoE)
(Mixture of Experts)
Many specialist sub-models. Only a few wake up per token. You get a huge brain at a small compute bill.
Definition
An architecture where a large model contains many specialised sub-networks (experts) and a router decides which small subset of experts processes each token, so most parameters stay inactive on any one forward pass.
Step 1