Even when it's "one big model" that "model" has usually been a moe for the last couple of years. Which is several smaller models in a trench coat. They each respond and the result you see is an amalgam.

I don't want to oversell my understanding of the topic, but I think at least the big corpo chatbots have the resources to run them entirely separately like that. I'm not sure how smaller models are compressed down or if they're able to staple some of these specialist models into the main ones to make the One Big Model That Can Do Everything. Apparently text-based models are capable of doing image now so maybe they can?