I suppose siglipv2 wasn't out yet when they were training this - I wonder if there will be an update to the multimodal models or pali-gemma which utilizes Siglip2. Aya Vision from Cohere utilized siglip 2 to great effect