What’s the maximum physical amount of intelligence we can fit into small models?

I was talking to my friend the other day, who is a really avid supporter of local models. He thinks we might be able to get something as intelligent (not smart in terms of how much it knows) as Claude Fable in inside a model which is like, 20 billion parameters, even if we’d get that in 2035.

While I don’t know much about small models since I’ve only recently come from the closed-source labs, is this even possible? I say no. And what is the minimum model size you think someone would need to host something as capable as Fable 5?