More parameters mean more patterns a model can hold, but memory and compute grow with them. Parameter count is therefore a performance figure and a hardware bill at the same time. For open models it decides whether the model can run on the hardware you already own.