If every layer prefix is a valid model, why do we still pick a size at deploy time?

I keep four checkpoints of the same family on disk: a 1.5B, an 8B, a 32B, and a 70B. Four training...