No, it would mean 13x times smaller models where training data is better curated.

Ironically I think this is a step towards Clip again, which Google already improved with SigLIP but now returns to introduce a global or "batch" perspective. At least that is how I read it with my very amateur understanding.