Understand physical common sense and generate appropriate embodied decisions. optimized for document-level optical character recognition, long-context vision-language understanding. build with hand-curated dataset for text-to-image models, providing significantly more detailed descriptions or captions of given images.