my intuition is that cross modality improvements do not really show up in practice, and especially not in small models