A 1B-parameter vision model just beat a 7B one on depth, frozen, single linear layer, zero fine-tuning.

@robbyant_brain's LingBot-Vision: 0.296 NYU-Depth v2 RMSE vs DINOv3-7B's 0.309. 7x smaller. Open weights.

The trick: train on boundaries, not just semantics.

Robbyant releases LingBot-Vision, the world's first spatially-native visual foundation model, a shift in the visual pre-training paradigm

Most vision foundation models are trained to become stable against visual changes, so they get very good at saying “cat,” “chair,” or “table” even when lighting or angle changes.

That is useful for recognition, but robotics needs something more annoying and physical: where the object starts, where it ends, where the thin cable is, where the transparent cup breaks depth sensing, and which boundary should stay stable as the robot moves.

LingBot-Vision is trained around boundaries.
It learns boundary-bearing visual tokens by itself, without human labels, without an external edge detector, and without starting from another pretrained backbone.

Those boundary tokens then become training targets for dense visual features, so the model learns scene structure at the patch level rather than only learning high-level semantics.

The result is a frozen visual backbone that does well on depth, segmentation, and video object tracking, including 0.296 RMSE on NYUv2 depth, ahead of 7B-parameter DINOv3 at 0.309, while using far fewer parameters.

🧵 1.