Nvidia Labs releases C-RADIOv4 distilled vision model for FiftyOne

Nvidia Labs has introduced C-RADIOv4, an agglomerative vision foundation model that distills SigLIP2, DINOv3 and SAM3 into a single architecture. It ships in H and SO400M variants and reaches 83.09% and 82.01% zero-shot accuracy on ImageNet-1K with far fewer parameters than its teachers.

• Zero-shot accuracy of 83.09% (H) and 82.01% (SO400M) on ImageNet-1K
• kNN accuracy of 86.59% and 85.76% on ImageNet-1K
• Segmentation of 55.20 and 55.14 mIoU on ADE20k
• C-RADIOv4-H rivals DINOv3-7B with far fewer parameters