The way to think of this is that (current) inferencing hardware has natural cut points of efficient scale (assuming some optimal architecture which at this point is almost always an MOE variant), so you always have an incentive to provide 3-4 model sizes (which nicely enough often means diff speed to answer also).

You are also rolling the training of your models by size and the lat est small model can sometimes beat an older model.

For a while we thought you just train the big model first then distill smaller models from it but since we've learnt the importance of on policy RL (basically the way a small model figures out a problem is very diff from a big model so it's best to RL optimize each model distinctly)