LingBot-Vision: 0.3B ViT-L matches DINOv3-7B on depth

Ant Group has released LingBot-Vision, a self-supervised DINO-family vision backbone available in four sizes—21M ViT-S, 86M ViT-B, 0.3B ViT-L and 1.1B ViT-g—under an Apache-2.0 license. Its key innovation is boundary-driven masking: a teacher network predicts where object boundaries lie and forces those tokens into the student's mask, preventing the model from solving reconstruction by simply copying flat background context. No labels, text supervision, or external edge detectors are used.

According to the reported benchmarks, the 1.1B flagship achieves the best NYUv2 depth RMSE in the comparison (0.296), beating DINOv3-7B (0.309) and V-JEPA 2 at 2B (0.307). More strikingly, the much smaller 0.3B ViT-L reaches 0.310—essentially matching the 7B model—despite having roughly 23x fewer parameters and being trained on only about a third of DINOv3's training images (161M).

The model isn't uniformly better: it trails DINOv3 on ImageNet classification at the larger scales (though the smaller B/S variants lead their respective classes), and larger models still win on KITTI. Since evaluations follow the standard DINOv3 frozen linear-probe protocol, the claims are easy to independently verify, and all figures remain self-reported pending outside reproduction.