Flow - 4x faster than typing
Mini toolsDINOv3
DINOv3 is Meta's third generation self supervised vision foundation model, scaling training to 7 billion parameters and 1.7 billion images while keeping backbones frozen at evaluation time. It builds on DINOv2 with a strong focus on dense image features, delivering state of the art results across object detection, semantic segmentation, depth estimation and video object tracking without task specific finetuning. Meta releases a full model suite spanning Vision Transformers from 21 million to 6.7 billion parameters plus efficient ConvNeXt backbones for on device deployment, including a satellite imagery backbone. Pre trained backbones can be adapted to new domains by training a lightweight adapter on a small amount of labeled data, making DINOv3 useful in domains such as satellite canopy height mapping, medical and organoid imaging, and robotics vision.
Image
Overview
DINOv3 is a self supervised computer vision model from Meta that produces universal image backbones without using human labeled data. Trained up to 7 billion parameters on 1.7 billion images, it generates high resolution dense visual features that power tasks like object detection, image segmentation, depth estimation and video object tracking without any finetuning.
About Meta Platforms
We're connecting people to what they care about, powering new, meaningful experiences, and advancing the state-of-the-art through open research and accessible tooling.
Industry:
Technology, Information and Media
Company Size:
78865
Location:
Menlo Park, California, US
Website:
ai.meta.com
Last updated: July 15, 2026
