DINOv3
Self-supervised vision foundation model trained on 1.7B images.
About
DINOv3 is Meta's third-generation self-supervised vision backbone: a 7B-parameter Vision Transformer that learns from 1.7 billion unlabeled images, with a new Gram Anchoring technique that keeps dense features stable at scale. For the first time, a single frozen backbone beats specialized models on dense tasks like segmentation, depth estimation, and tracking without fine-tuning, and it ships in sizes from 21M to 7B.
It is an encoder, not a chatbot — you still need a task head or a VLM on top. Weights are available under a Meta commercial license rather than a standard open-source license.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- First frozen backbone to beat specialized dense-prediction solutions across multiple tasks (Meta-reported)
- Dense features at 1024x1024 resolution
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Meta Research release (Aug 2025); 7B model plus distilled variants 21M-1B with training code public; Adopted quickly as a frozen feature extractor in vision research and geospatial/earth-observation tooling; Self-supervised recipe published, but training data (1.7B images, LVD-168M) not released; Distributed under a Meta commercial license — usable commercially, but not a standard OSS license. License: Meta commercial license (custom) — free for commercial use but not OSI-open.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 7.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Meta Research release (Aug 2025); 7B model plus distilled variants 21M-1B with training code public
- · Adopted quickly as a frozen feature extractor in vision research and geospatial/earth-observation tooling
- · Self-supervised recipe published, but training data (1.7B images, LVD-168M) not released
- · Distributed under a Meta commercial license — usable commercially, but not a standard OSS license