DINO (DINOv2, DINOv3)

weights

open(every DINOv3 checkpoint on the Hub behind a manual access request)

data

documented-not-released(LVD-1689M, curated from public Instagram posts, and SAT-493M satellite imagery

code

open(full pretraining pipeline, configs and distillation recipes)

license

DINOv3-License(code and weights

DINOv3's code and weights come under Meta's own DINOv3 License, which permits commercial use but bars military, weapons and trade-controlled uses and lets Meta change the terms; downloads need an approved access request. The full pretraining code is published, but the 1.7-billion-image training set drawn from Instagram is not. The earlier DINOv2 was Apache-2.0, but the current release governs. The DINOv3 License caps neither who may use the model nor at what scale, so its tier is permissive_non_osi.

• https://huggingface.co/api/models?author=facebook&search=dinov3&sort=downloads&limit=100&expand[]=downloads&expand[]=cardData&expand[]=gated recorded 2026-09-27

  All thirteen facebook/dinov3-* checkpoints are gated "manual" and carry the dinov3-license (the chmv2 head without a license tag); none carries another license.

• https://raw.githubusercontent.com/facebookresearch/dinov3/main/LICENSE.md recorded 2026-09-27

  LICENSE.md is the "DINOv3 License", "Last Updated: August 19, 2025"; 1.a grants "a non-exclusive, worldwide, non-transferable and royalty-free limited license ... to use, reproduce, distribute, copy, create derivative works of, and make modifications to the DINO Materials"; 1.b.v bars use "related to military or warfare purposes, nuclear industries or applications, espionage, or the development or use of guns or illegal weapons". No non-commercial clause.

• https://raw.githubusercontent.com/facebookresearch/dinov3/main/MODEL_CARD.md recorded 2026-09-27

  Model card: "Web dataset (LVD-1689M): a curated dataset of 1,689 millions of images extracted from a large data pool of 17 billions web images collected from public posts on Instagram"; no release of the dataset.

• https://raw.githubusercontent.com/facebookresearch/dinov3/main/README.md recorded 2026-09-27

  README License section: "DINOv3 code and model weights are released under the DINOv3 License."; Training section documents pretraining, gram anchoring and high-resolution adaptation for ViT-7B/16.

• https://ungh.cc/repos/facebookresearch/dinov3/files/main recorded 2026-09-27

  Repository tree includes dinov3/train/train.py, ssl_meta_arch.py, multidist_meta_arch.py and dinov3/configs/train/dinov3_vit7b16_pretrain.yaml, gram_anchor and high_res_adapt configs.

Hugging Face downloads over the trailing 30 days for the three most-used DINOv3 checkpoints, which are the ones declared. The other DINOv3 checkpoints and the older, more downloaded DINOv2 checkpoints would raise the total without changing its order of magnitude.

• https://huggingface.co/api/models?author=facebook&search=dinov2&sort=downloads&limit=100&expand[]=downloads&expand[]=cardData&expand[]=gated recorded 2026-09-27

  The facebook/dinov2-* checkpoints, ungated and tagged apache-2.0, sum to about 7.16 million downloads in the trailing 30 days.

• https://huggingface.co/api/models?author=facebook&search=dinov3&sort=downloads&limit=100&expand[]=downloads&expand[]=cardData&expand[]=gated recorded 2026-09-27

  dinov3-vitl16-pretrain-lvd1689m 619,462, dinov3-vitb16-pretrain-lvd1689m 584,214 and dinov3-vits16-pretrain-lvd1689m 427,912 downloads in the trailing 30 days (1,631,588 together).

DINOv3 is best known as a backbone that other models are built on, but Meta also ships ready heads, and its dino.txt variant classifies and segments whatever labels the user types with no training, the same kind of open-vocabulary capability that places Ultralytics here. Its detector and segmentor heads cover fixed benchmark label sets.

• https://raw.githubusercontent.com/facebookresearch/dinov3/main/README.md recorded 2026-09-27

  README sections: "Pretrained heads - Image classification", "Pretrained heads - Depther trained on SYNTHMIX dataset", "Pretrained heads - Detector trained on COCO2017 dataset", "Pretrained heads - Segmentor trained on ADE20K dataset", "Pretrained heads - Zero-shot tasks with dino.txt"; notebook "Zero-shot segmentation with DINOv3-based dino.txt: compute the open-vocabulary segmentation results".