If it's a new pretrain, the token embeddings could be wider - you can pack more info into a token making it's way through the system.

Like Chinese versus English - you need fewer Chinese characters to say something than if you write that in English.

So this model internally could be thinking in much more expressive embeddings.