What is DINOv3?
DINOv3 is a family of image encoders from Meta FAIR, released on 14 August 2025. An encoder turns a photo into a grid of numbers, one per small patch, that other models can build on. DINOv3 learned to do this from 1.7 billion unlabelled web images, with no captions or human tags, which Meta says is 12 times the data and 7 times the model size of DINOv2. The flagship is a 7B-parameter vision transformer (6,716M parameters); Meta distilled it into ViT models of 21M, 29M, 86M, 300M and 840M parameters and ConvNeXt models of 29M to 198M for smaller hardware.
The selling point is that the features work frozen. Plug a light head on top and the same backbone does classification, object detection, semantic segmentation, depth estimation and video tracking without fine-tuning the encoder. A new training trick Meta calls Gram anchoring stops the per-patch features turning noisy over long training runs, a known weakness of the approach. Separate ViT-L and 7B backbones trained on 493 million MAXAR satellite images target mapping work, and the dino.txt add-on aligns features with text for zero-shot segmentation.
The weights are not Apache or MIT. They ship under the DINOv3 License, a Meta licence that allows commercial use but bars military, ITAR-related, nuclear and espionage uses and reverse engineering, and lets Meta change the terms. Downloads on Hugging Face need a request that is approved by hand.
Is DINOv3 worth it?
DINOv3 is the strongest general-purpose image backbone you can download, and it is the first, in Meta's words, where a single frozen encoder beats specialised systems on dense tasks. With the backbone frozen and a detection head on top it reaches 66.1 mAP on COCO, and with a segmentation head 63.0 mIoU on ADE20k, both state of the art when published. With only a linear layer on top, ADE20k segmentation is 55.9 mIoU, more than 6 points ahead of other self-supervised models and 13 ahead of the weakly supervised baselines. These are Meta's own numbers across more than 60 benchmarks.
The practical gain is that one forward pass serves several tasks. A drone or factory camera can run the encoder once and feed the same features to a detector, a segmenter and a depth head, which Meta pitches for edge devices. The distilled family matters as much as the flagship: the 300M ViT-L is the most downloaded checkpoint, at about 700,000 downloads in the last 30 days on Hugging Face, and Meta says the distilled ViT-B and ViT-L beat comparable CLIP-based models.
Satellite work is where it already earns its keep. The World Resources Institute uses it to map tree canopy height; in one region of Kenya average error fell from 4.1 metres with DINOv2 to 1.2 metres, according to Meta. Its Canopy Height Maps v2 model, released in March 2026, is built on DINOv3 and loads in Transformers.
The licence is the cost. DINOv2 is under Apache 2.0; DINOv3 uses Meta's own terms, with a manually approved download, a ban on military, nuclear and ITAR-related uses, a requirement to credit it in publications, and a clause letting Meta modify the agreement with immediate effect. DINOv3 is not a text model, either: it does not caption or answer questions, and zero-shot work by name needs the separate dino.txt head, where CLIP-style models such as SigLIP 2 remain the simpler choice.
Where it earns its place
A remote-sensing analyst at a forestry or climate non-profit
Estimate tree canopy height and spot deforestation from satellite imagery
Start from the SAT-493M backbone or Meta's CHMv2 canopy-height model in Transformers. WRI reports error falling from 4.1 m to 1.2 m in one Kenyan region against DINOv2, according to Meta.
A perception engineer at a warehouse-robotics company running Jetson-class hardware
Run detection, segmentation and depth from one camera on a tight compute budget
Distilled backbones such as the 21M ViT-S or 29M ConvNeXt Tiny run once per frame and feed several light heads, instead of three separate networks. Accuracy drops against the 7B model, so benchmark the size against your own scenes.
An ML lead at a 15-person retail-analytics start-up
Find visually similar products across a catalogue of 2 million photos
Embed every image with the 300M ViT-L and index the vectors in an existing vector database; no labels are needed. Check the DINOv3 License with counsel first, since its terms are Meta's own and can change.
Before you commit
- Setup
- Request access on Hugging Face and wait for manual approval, then load with Transformers' pipeline or AutoModel (4.56+) or timm 1.0.20+. Direct downloads come by e-mailed URLs after Meta's form; the README says to use wget rather than a browser.
- Integrations
- Transformers, timm and PyTorch Hub; pretrained heads for COCO detection, ADE20k segmentation and NYUv2 depth; dino.txt for zero-shot text-to-image matching; notebooks for video segmentation tracking and zero-shot segmentation.
- Data and privacy
- Self-hosted, so images stay on your hardware. The DINOv3 License allows commercial use and derivatives but bans ITAR-related, military, nuclear, espionage and weapons uses and reverse engineering, requires crediting in publications, and lets Meta modify the terms.
- Cost at scale
- The 7B ViT holds about 13.4 GB of weights at 16-bit, by our arithmetic, so one 24 GB or larger GPU serves it; the distilled 21M to 840M models run on consumer GPUs and edge devices. Meta estimates the whole project used about 9 million GPU hours, so retraining from scratch is not realistic.
- Limits
- Image-only; no text output and no video model of its own. Benchmarks are self-reported; the strongest results use the 7B model with task heads, and smaller distilled models score lower.
What’s good
- Frozen backbone reaches 66.1 mAP on COCO detection and 63.0 mIoU on ADE20k (Meta)
- Sizes from 21M to 7B parameters, in ViT and ConvNeXt variants
- Dedicated satellite backbones trained on 493M MAXAR images
- Supported in Transformers 4.56+ and timm 1.0.20+
What’s not
- DINOv3 License, not Apache 2.0 like DINOv2, with use bans and changeable terms
- Hugging Face access is approved by hand
- No text understanding without the separate dino.txt head
- 7B flagship needs a data-centre GPU; most teams will use a distilled model
Best for
Computer-vision teams that want one frozen image encoder to feed detection, segmentation, depth or retrieval heads, including on satellite and aerial imagery.
DINOv3 alternatives
DINOv3: common questions
- Is DINOv3 worth it?
- DINOv3 is the vision backbone a computer-vision team should start from in 2026, provided legal is happy with Meta's licence. We rate it 4.6 out of 5.
- How much does DINOv3 cost?
- Free under the DINOv3 License after a manually approved request; commercial use allowed, with military, ITAR, nuclear, espionage and reverse-engineering restrictions. No hosted API from Meta.
- What are the downsides of DINOv3?
- DINOv3 License, not Apache 2.0 like DINOv2, with use bans and changeable terms; hugging Face access is approved by hand; no text understanding without the separate dino.txt head; 7B flagship needs a data-centre GPU; most teams will use a distilled model.
- Who is DINOv3 best for?
- Computer-vision teams that want one frozen image encoder to feed detection, segmentation, depth or retrieval heads, including on satellite and aerial imagery.
- How much does DINOv3 cost for a team or in production?
- The 7B ViT holds about 13.4 GB of weights at 16-bit, by our arithmetic, so one 24 GB or larger GPU serves it; the distilled 21M to 840M models run on consumer GPUs and edge devices. Meta estimates the whole project used about 9 million GPU hours, so retraining from scratch is not realistic.
- Is DINOv3 safe for company data?
- Self-hosted, so images stay on your hardware. The DINOv3 License allows commercial use and derivatives but bans ITAR-related, military, nuclear, espionage and weapons uses and reverse engineering, requires crediting in publications, and lets Meta modify the terms.
- What are the best DINOv3 alternatives?
- The closest alternatives we have reviewed are V-JEPA 2 and V-JEPA 2.1, SAM 3 (Segment Anything Model 3), Sapiens2 and Qwen3-VL-Embedding.
Answers checked Nov 9, 2026.
Make DINOv3? Show this rating on your site
