DINOv3
7B-param self-supervised vision backbone trained on 1.7B unlabeled images; frozen features beat specialised models on dense tasks.
Scaled DINO 6x in parameters and ~12x in data; claimed first frozen backbone to outperform specialised solutions on detection, segmentation and depth, and shipped a satellite-imagery variant (tree canopy height error cut from 4.1 m to 1.2 m in Kenya).
- Date
- Thursday, 14 August 2025
- Lab
- Meta
- Kind
- open-weights
- Access
- open weights (restricted license)
Figures
| Measure | Value | Measured by |
|---|---|---|
| Scale | 7B params, 1.7B images ViT-B/L and ConvNeXt distilled variants released | company |
Released under a custom commercial-use license (not Apache 2.0).
Sources
This record was checked against its sources on 6 October 2026. How we check