general family · filed under model architecture
Token-level expert routing
A router sends each token to a subset of experts, but the evidence does not specify the selection algorithm.
Also called token routing through 8 of 288 experts, routes each token through 8 of 288 experts.
- sources
- 3
- models
- 3
- labs adopt it
- 3
- strongest
- core
How sources treat it
One count per evidence span, weakest treatment to strongest.
core 3
Documented in
Evidence
3 spans quoted from the sources, strongest treatment first.
MoEs are fast and efficient because a router sends each token to just a few of its many experts, so only a fraction of the model’s parameters run per token.
coremodel architecturein Nemotron 3.5 LightningNVIDIA
each token is routed to 6 of 256 experts, plus 2 shared experts active on every token
coremodel architecturein InklingThinking Machines Lab
routes each token through 8 of 288 experts
coremodel architecturein GLM-5.3-FlashZ.ai
Filed alongside
Other methods under model architecture :: channel mixer :: mixture of experts :: expert routing.
Top-8 expert routingAnticipatory RoutingTop-4 expert routingDP-aware routingFixed Top-k routing with frozen biasHash routingKeep RoutingLatent-space routingLoss-spike-triggered Anticipatory RoutingRouted expertsToken-choice routingToken-choice routing with softplus gatingTop-6 expert routingTop-k expert routing with softmax over selected experts