Ant Group released LingBot-Vision: DINO-family vision backbones in 4 sizes, and the 0.3B ViT-L matches DINOv3-7B on NYUv2 depth with ~23x fewer params
蚂蚁集团发布了 LingBot-Vision，这是一个包含四种尺寸的 DINO 家族视觉骨干网络。该模型采用边界驱动的自监督掩码机制，教师模型预测物体边界，学生模型则被迫在这些区域进行重建。LingBot-Vision 的 0.3B ViT-L 模型在 NYUv2 深度估计任务上表现与 DINOv3-7B 相当，但参数量减少了约 23 倍。…
Weights, all 4 sizes, Apache-2.0 (ViT-S / ViT-B / ViT-L / ViT-g): https://huggingface.co/collections/robbyant/lingbot-vision Code: https://github.com/robbyant/lingbot-vision Project page: https://technology.robbyant.com/lingbot-vision
Self-supervised DINO-family backbone, but the masking is boundary-driven: the teacher predicts where object boundaries are and those tokens get forced into the student's mask, so it can't solve reconstruction by copying flat context. No labels, no text supervision, no external edge detector.
The screenshot is Table 2. The 1.1B flagship gets the best NYUv2 RMSE in the whole comparison (0.296 vs 0.309 for DINOv3-7B, with V-JEPA 2.1 at 2B in between at 0.307). The 0.3B ViT-L number (0.310, basically the 7B's 0.309) is from Table 5. Sizes per the report: 1.1B ViT-g, 0.3B ViT-L, 86M ViT-B and a 21M ViT-S. ViT-L is ~0.6GB in fp16. The repos ship a small custom loader (lbot_vision_infer) that returns the frozen backbone for feature extraction, and dense features (depth/segmentation probes, tracking) are the point, not chat.
Where it loses, same tables: ImageNet classification trails DINOv3 at the flagship and L scales (the B/S students actually lead their class on linear probe), and KITTI favors the bigger models (both the 7B and the 2B V-JEPA beat it there). Trained on 161M images, less than 1/3 of DINOv3's training samples.
All numbers are self-reported. The eval protocol is the standard DINOv3 frozen linear-probe recipe, so it's cheap to check independently. Given how the Ling-1T threads went, I'd treat the deltas as unverified until someone reruns the probe.