Meta Platforms has released DINOv3, a new family of computer-vision foundation models trained with self-supervised learning. The company says a single, "frozen" DINOv3 backbone surpasses specialized systems on long-standing dense-prediction benchmarks while operating across a range of image resolutions. Model checkpoints in multiple sizes—including a top configuration of roughly 7 billion parameters—are now available on the open-source repository Hugging Face under a permissive commercial license. Meta’s Fundamental AI Research (FAIR) group also published the accompanying technical paper to encourage external validation and downstream development.