active parameters
The portion of a model's parameters actually switched on for any single word it produces, which in mixture-of-experts designs is far smaller than the total.
Parameters are the adjustable numbers inside a model, and headline sizes quote all of them. Active parameters quote only the ones actually used for a given word. In a dense model those are the same number. In a mixture-of-experts design, where the model is split into specialised sections and a router picks a few for each token, they are very different: a model advertised at 400 billion parameters might switch on 30 billion at a time.
The distinction matters for two practical reasons. Cost and speed track the active count, because that is the arithmetic actually being done, which is why mixture-of-experts models can be much cheaper to run than their headline size implies. Memory, on the other hand, tracks the total, because all the parameters have to be loaded somewhere even if most sit idle. That is the trap when running one at home: a model may be fast enough and still not fit.