Just to expand your lower bar, some people suspect it is not much bigger than GPT 3.5 @ 175B parameters, but this is also speculation. I don't think there has been anything official.
Given how PALM2 had less params than the first one (540B) I think that makes sense.
Can't be less though, the latency difference between them implies that GPT-4 is significantly larger. This is unlike with PaLM, where latency decreased.