Yeah but they're all trained together along with their gate tensor. Evaluating a moderate length prompt pulls in most/all of them.