Technology

DINOv3-ConvNext

A family of highly efficient, self-supervised vision models distilled from Meta's 7-billion-parameter DINOv3 teacher into lightweight ConvNeXt architectures for fast, on-device deployment.

DINOv3-ConvNeXt bridges the gap between massive foundation models and edge deployment by distilling the visual representations of Meta's 7B-parameter Vision Transformer (ViT) into efficient convolutional backbones (ranging from 29M to 198M parameters). Trained on a curated dataset of 1.7 billion unlabeled images, these models bypass the need for costly fine-tuning while delivering state-of-the-art performance on dense prediction tasks like object detection, semantic segmentation, and depth estimation. By swapping complex ViT architectures for streamlined ConvNeXt variants (Tiny, Small, Medium, and Large), developers gain access to powerful, high-resolution features optimized specifically for memory-constrained and speed-critical environments.