Active parameters
The subset of an MoE model's parameters involved in producing any single token.
Written as the second figure in names like 35B-A3B: 35 billion total parameters, about 3 billion active per token. Total parameters determine how much memory you need; active parameters determine how fast it generates. A model's headline parameter count is the total, so quoting it as though it governed speed misrepresents every MoE model.