DINOv3

DINOv3 is a family of self-supervised vision foundation models developed by Meta AI, unveiled on August 14, 2025. The models are trained without labels on 1.7 billion images (the LVD-1689M dataset), with the flagship ViT-7B variant containing about 6.7 billion parameters.

Meta also released distilled variants (ViT-S, ViT-S+, ViT-B, ViT-L, ViT-H+ and ConvNeXt T/S/B/L backbones) and a satellite variant trained on the SAT-493M dataset. A key technique is Gram anchoring, which addresses the degradation of dense feature maps during long training schedules.

The models achieve state-of-the-art results on object detection, semantic segmentation and depth estimation without fine-tuning the backbone. Weights and code are released under the DINOv3 License via GitHub (facebookresearch/dinov3) and Hugging Face.