How many weights the model carries.
Parameters are the learned numbers inside a model. Count them and you get a rough sense of capacity, memory footprint and cost to serve.
Bigger stopped meaning better some time ago: data quality, training length and post-training now separate models of identical size, and mixture-of-experts designs activate only a fraction of their parameters per token.
For deployment the number that matters is active parameters and the memory they need, because that decides which hardware can host the model at all.