Ant Group Open-Sources LingBot-Vision: A Boundary-Centric Vision Model
Ant Group's open-source LingBot-Vision turns boundary predictions into a self-supervised training signal, beating the 7B DINOv3 with far fewer parameters.
Ant Group's Robbyant unit has open-sourced LingBot-Vision, a 1.1B-parameter ViT-g/16 that treats boundaries as a training signal rather than a downstream output. The model is trained purely through self-supervision, with no labels, external edge detectors, or pretrained backbones involved. Boundary tokens discovered by the teacher network are forced into the student's masked set, turning the least redundant regions into the hardest prediction targets.
Boundary fields are reparameterized as categorical distributions over 32 bins per channel, avoiding regression collapse while inheriting DINO-style centering and sharpening. The uniform distribution over bins doubles as an a-contrario null hypothesis, letting an NFA test validate every segment so unsupported structure never becomes part of the learning signal.
The results are notable: on NYUv2 linear probing, the model beats the 7B-parameter DINOv3 while using roughly 7x fewer parameters and a training corpus an order of magnitude smaller. Even a distilled 0.3B student matches DINOv3's performance. Swapping the encoder into LingBot-Depth 2.0 sets state-of-the-art results across 14 depth-completion benchmarks, with the gap widening as training data scales.
For engineers, the takeaway is clear: embedding boundary awareness directly into pretraining enables efficient, verifiable foundation models for dense spatial perception with a fraction of the usual compute and data.