[torginus]

It's just how LLMs work - these are three completely separate models trained in parallel, with different numbers of parameters and using different amounts of compute.

They could hide this behind a harness that picks the correct model for you, but devs don't seem to like that.

There's also the 'effort' slider, which I guess how many experts in the MoE are evaluated and how long reasoning chains are allowed to go on, which is the 'smooth' scaling you are thinking of.