Cosmos 3: Omnimodal World Models for Physical AI

Benchmark Model Rank Results
visual-question-answering-vqa-on-realworldqaCosmos3-Super#14Accuracy: 79.2
object-counting-on-countbenchCosmos3-Super#15Accuracy: 89.1
video-question-answering-on-mvbenchCosmos3-Super#15Avg.: 74.5
visual-question-answering-on-docvqa-testCosmos3-Super#39ANLS: 0.904
multimodal-reasoning-on-mmmu-proCosmos3-Super#42Accuracy: 48.1
instruction-following-on-ifbenchCosmos3-Super#43Accuracy: 37.0