Meta releases DINOv3

Meta released DINOv3, a computer-vision model trained without human-labelled data on 1.7 billion images, scaled up substantially from its predecessor DINOv2 in both model size — a 7-billion-parameter vision transformer — and training data. The self-supervised approach lets the model learn general visual representations by predicting relationships between different views of the same unlabelled image, rather than being trained on a task-specific dataset of labelled examples.

Meta reported that the resulting frozen backbone — used without further fine-tuning — matched or outperformed specialised, task-trained networks on object detection, semantic segmentation and video tracking benchmarks. That result was notable because vision models had typically needed labelled training data and task-specific fine-tuning to reach competitive performance on those benchmarks; a single general-purpose backbone doing comparably well without either suggested self-supervised pretraining had closed most of that gap for vision, following a path text-based language models had taken earlier.

Meta released DINOv3 openly under a commercial licence, alongside a family of smaller models distilled from the full 7B backbone — including both vision-transformer and ConvNeXt variants — to suit different compute budgets, and a version trained specifically on satellite imagery. As with Meta’s earlier open releases, the aim was to seed adoption of Meta’s models as an infrastructure layer for other companies’ vision products, extending the open-versus-closed contest that had mostly played out in language models into computer vision.