Image models are much smaller than language models, so they put some extra language features on an image model here and then it doesn't need to be that large.