Overview
DINOv3 is a family of self-supervised vision foundation models from Meta’s FAIR research lab, released in August 2025 with a technical report (arXiv 2508.10104). It turns images into general-purpose visual features without needing any labeled data for training. Meta scaled the approach to a 7-billion-parameter vision transformer trained on 1.7 billion web images, then distilled it into smaller models that are practical to deploy. Meta says it is the first time a single frozen vision backbone beats specialized systems on long-standing dense tasks such as object detection and semantic segmentation.
How it works
DINOv3 continues the DINO line of self-distillation: the model learns by making its own views of an image agree, so no captions or labels are needed. The report adds “Gram anchoring”, a technique that stops the fine-grained, per-patch feature maps from degrading over long training runs, a known problem when scaling this kind of training. Post-training steps then make the models more flexible across image resolutions and model sizes, and align them with text through the dino.txt method for zero-shot tasks.
The released suite, per the GitHub repository, has 12 backbones with a patch size of 16:
- Web images (LVD-1689M): ViT-7B (6,716M parameters) plus distilled ViT-S (21M), ViT-S+ (29M), ViT-B (86M), ViT-L (300M) and ViT-H+ (840M).
- ConvNeXt Tiny, Small, Base and Large (29M to 198M) for resource-constrained use.
- Satellite imagery (SAT-493M, built from Maxar imagery): ViT-L and ViT-7B.
Meta also published task heads for classification, depth estimation, COCO object detection and ADE20K segmentation on top of the 7B backbone.
Licence and running it
Code and weights use the DINOv3 License, a custom Meta licence that grants a royalty-free right to use, modify and redistribute the models, including commercially, but with conditions such as passing the licence on and acknowledging DINOv3 in published research. Weights are gated: you request access through Meta’s download page or the Hugging Face collection. The backbones load with PyTorch Hub, with Hugging Face Transformers from version 4.56.0 and with the timm library from version 1.0.20.
Strengths and limits
Meta reports that DINOv3 matches or beats SigLIP 2 and Perception Encoder on many classification benchmarks while clearly leading on dense prediction. One real-world example from Meta’s announcement: the World Resources Institute used the satellite models to cut the average error of tree-canopy height estimates in a region of Kenya from 4.1 meters to 1.2 meters, compared with DINOv2. The models output features, not finished answers, so a task head is still needed, the 7B backbone is costly to run, and the licence is not an OSI-approved open-source licence.
Alternatives
Its predecessor DINOv2 is lighter and widely supported. For features aligned with text, see CLIP ViT-L/14 or SigLIP; for promptable segmentation, Meta’s SAM 3. Our explainer on how self-driving cars see and our overview of AI in space exploration, which covers satellite image analysis, show where such vision backbones are applied.
Specifications are taken from the GitHub repository, the technical report and Meta’s announcement. Benchmark results depend on evaluation settings; see how AI benchmarks work.