Has anyone tried/does it make any sense to use a small model to get a first opinion, and then if it is unsure what token to emit next run a bigger model? It see

Has anyone tried/does it make any sense to use a small model to get a first opinion, and then if it is unsure what token to emit next run a bigger model? It seems like a lot of tokens ought to be very obvious and not require as much computation. E.g. most words are split into multiple tokens, but most of the time in context after the first token for a word only one follow-on token makes sense.