Meta ships DINOv3 behind an access gate under its own license. Ant's Robbyant just shipped a full vision backbone family under Apache-2.0. What happens when perception goes free and small?
Robbyant, an embodied AI company under Ant Group, dropped LingBot-Vision, a self-supervised vision backbone family ranging from 21M to 1. Their stated mission is building one brain for all robots.1B parameters, all Apache-2.0 on HuggingFace and GitHub. The release includes code, pretrained weights, and a project page with interactive point-cloud comparisons across eight methods. This is not a paper drop with a promise of weights later. The weights are live now.
The architecture sits in the DINO lineage but with a twist they call masked boundary modeling. The teacher predicts a dense boundary field online, and tokens that carry boundaries are forced into the student's mask. Boundary fields get recast as per-pixel categorical distributions to keep self-distillation stable, and decoded segments pass an a-contrario validation test. No labels, no text supervision, no external edge detector. They trained on 161M curated images, which they report is less than one third of DINOv3's training samples.
On their self-reported numbers using a frozen linear-probe protocol, the 1.1B ViT-g flagship hits 0.296 RMSE on NYUv2 depth, which they place ahead of DINOv3-7B at 0.309. The distilled ViT-L at 0.310 basically matches that DINOv3-7B score at about one twenty-third the parameters. But they also show losses. On KITTI depth, LingBot scores 2.552 while DINOv3-7B hits 2.346. On ImageNet linear probing, the flagship trails DINOv3-7B, though the ViT-B and ViT-S variants reportedly lead their size classes. For segmentation, they report being roughly on par with distilled DINOv3 ViT-H+ across ADE20K, Cityscapes, and VOC, with some swaps in either direction.
The downstream product is LingBot-Depth 2.0, a depth-completion model that fills in glass, mirrors, and transparent surfaces where RGB-D sensors return nothing. Those weights are not released. Only the four vision backbones are open. You also need their custom inference library rather than plain transformers or timm. ViT-L is about 0.6GB in fp16.
Perception, not the chat layer, is what robots actually run on. It is the raw spatial understanding that turns sensor input into something a system can act on. When that layer becomes small enough to run on edge hardware and free enough to modify without license friction, the stack above it shifts. A 21M parameter variant that reportedly leads its size class changes what you can embed in a cheap camera module.
The interesting contrast is release strategy, not geography. Meta released DINOv3 under its own gated license, not an OSI one. Robbyant released four sizes, Apache-2.0, no gate. If dense spatial tasks keep trending open while generative video stays closed and API-gated, do we end up with a split world where physical AI runs on open perception and digital AI runs on closed generation? Or does the pressure eventually force the closed labs to release vision weights too?