A conventional model runs every one of its parameters for every request. A mixture-of-experts model instead divides those parameters into many smaller sub-networks, called experts, and uses a routing layer to send each piece of input to a small number of them. A model with a trillion total parameters might activate only thirty billion of them to answer any given question.
That is why headline parameter counts have become misleading. Total parameters determine how much memory the model needs to be loaded at all, while active parameters determine how much compute each request consumes. A large mixture-of-experts model can be cheap to run and still expensive to host, which is exactly the combination that makes memory capacity, rather than raw processing speed, the binding constraint on running big models locally.
For anyone costing an AI deployment, ask for both numbers. A vendor quoting total parameters is describing the hardware you need to buy. A vendor quoting speed is describing the active subset. Neither figure on its own tells you what the system will cost to run at your volume.