Why is the first AI generation after switching models slower?

The local runtime may need to load the model into memory before generation starts. After that, repeated requests on the same model are usually faster.