DINOv3 ViT-7B/16 (LVD-1689M)
Release in the DINOv3 family · version ViT-7B/16, pretrain LVD-1689M
Maintained by Meta (FAIR)13
The largest DINOv3 backbone: a 6.7-billion-parameter Vision Transformer with 16-pixel patches, trained by Meta with self-supervised learning (a DINO self-distillation loss and an iBOT masked-image-modeling loss) on LVD-1689M. Training ran in three stages: pretraining, Gram anchoring, and high-resolution adaptation.135
- Model hub: Model card (Hugging Face, gated) (external site: huggingface.co)
- Repository: DINOv3 repository (external site: github.com)
- License: DINOv3 License (external site: github.com)
- Paper: DINOv3 paper (arXiv 2508.10104) (external site: arxiv.org)
Availability and license
Overall availability
Access requires a request and acceptance of the DINOv3 License. Through Meta's download page, approved users receive download URLs by email; the Hugging Face repository is gated with manual approval and asks users to share contact information.312
Availability is separate from permission: read the license before using or redistributing.
Custom Meta license (last updated August 19, 2025); the README applies it to both the code and the model weights. It grants a non-exclusive, worldwide, non-transferable, royalty-free license to use, reproduce, modify, and distribute, and Meta's announcement describes it as a commercial license. Redistributions must include the agreement, and published research must acknowledge use of the DINO Materials. Users must comply with trade controls, may not be sanctions targets, and may not use the materials for ITAR-regulated activities or prohibited end uses including military or warfare, nuclear, espionage, or weapons applications. It also forbids reverse engineering, ends for anyone who brings IP litigation against Meta over the materials, is governed by California law, and may be modified by Meta.436
Model-disclosure tier
The weights can be obtained only by request, with approval, or by some users — for example a gated download that the publisher reviews. Not counted as open-weight.
Public materials checklist
| Item | Status | Notes and evidence |
|---|---|---|
| WeightsCan the general public download the model parameters for this release? | Partial | Approval-gated on Hugging Face (manual) and on Meta's download page.23 |
| Inference codeIs code for running the model published? | Public | The repository loads backbones and heads through PyTorch Hub; the model card documents use with Hugging Face Transformers, which supports DINOv3 from version 4.56.0.31 |
| Training codeIs the code used to train the model published? | Public | The repository publishes the training code with commands and configuration files for the exact ViT-7B/16 setup (pretraining, Gram anchoring, and high-resolution adaptation) and for multi-distillation.3 |
| Training-data informationPublic = the training data itself can be obtained. Partial = composition or sources are documented without full access. | Partial | The model card describes LVD-1689M as 1,689 million images curated from a pool of 17 billion web images from public Instagram posts. The README states that ViT-7B/16 was trained on a private dataset, which is not released.13 |
| Training recipeAre the training configuration and procedure documented in enough detail to follow? | Public | Stage-by-stage configuration files for the ViT-7B/16 run are published, and the paper and model card describe the objectives and training procedure.315 |
| Evaluation materialsPublic = evaluation code or prompts that let others re-run the evaluations are published. Partial = results only. | Public | The repository publishes evaluation code and commands, including ImageNet-1k logistic regression, ADE20K linear segmentation, and NYUv2 linear depth probing.3 |
What it is useful for
The model card lists use of frozen features for image classification with k-NN classifiers, retrieval, depth estimation and semantic segmentation with linear layers, unsupervised object discovery, and video segmentation tracking. It also reports performance differences across income levels and regions in a fairness analysis.1
Run and use notes
Organization context
Other releases in the DINOv3 family
No other releases in this family have been assessed.
U.S. eligibility
Eligible · basis: U.S. headquarters
The model card names Meta AI as the developer. Meta Platforms, Inc. has its principal executive offices in Menlo Park, California, per its Form 10-K. The DINOv3 License names Meta Platforms Ireland Limited as licensor for EEA and Swiss users and Meta Platforms, Inc. for everyone else.147
Sources
This listing is not an endorsement, a safety assessment, or a federal approval.