What 100 million parameters means

When a model has 100 million parameters, training was allowed to change 100 million adjustable numbers: the weights on the connections and the biases on the neurons.

They are not 100 million written rules. One number says almost nothing by itself. Useful behaviour comes from all the values working together in the right places.

Finding those values can require enormous amounts of data, calculation, and repeated correction. Copying the finished parameter file is much easier.

The parameter count tells you how many learned settings the model contains. The saved file preserves the result of training them.