DINOv3

No labels went into DINOv3: Meta trained its third-generation vision foundation models with self-supervised learning on LVD-1689M, a curated set of 1.7 billion web images, plus a 493-million-image satellite corpus for geospatial variants. The result is a family of backbones from a 21M-parameter ViT-S up to a 6.7B ViT-7B, with ConvNeXt distillations from 29M to 198M for tighter deployment budgets, all producing high-resolution dense features that transfer to detection, segmentation, depth estimation, 3D correspondence, and video tracking without fine-tuning the backbone. A Gram-matrix anchoring technique keeps patch-level features clean during long training runs, the key fix over DINOv2, and frozen-backbone probes beat weakly supervised rivals on many dense prediction benchmarks. Checkpoints are distributed under the custom DINOv3 License, which allows commercial use with conditions rather than following a standard OSI license, through a request flow on Meta's site and Hugging Face. The repo provides PyTorch code (2.7.1 or newer), starter notebooks, and a text-alignment recipe, with a GPU expected for practical feature extraction.