Ant Group’s LingBot-Vision: A Game Changer?

Ant Group’s Robbyant team just open-sourced LingBot-Vision, a 1.1B model that fundamentally flips the script on spatial perception. This research marks a massive shift in how we approach vision models.

Usually, vision models treat object boundaries as an afterthought—something a task head tries to guess after the main training is done. But LingBot-Vision uses "boundary-forcing masking." Essentially, the model identifies boundaries during self-supervised training and forces itself to focus on those hardest-to-predict tokens. It’s incredibly clever because it doesn't need human labels or external edge detectors to learn what a boundary looks like.

The performance numbers are what really caught my eye. Their 0.3B distilled student model is actually matching the performance of the 7B DINOv3 on NYUv2 benchmarks. We are talking about a model that is roughly 23x smaller while delivering comparable results. It even crushed depth-completion benchmarks when used in their LingBot-Depth 2.0 setup.

In my opinion, this is a huge win for efficient AI. We’ve been stuck in a "bigger is better" loop with massive parameter counts, but this proves that if you refine the training signal—specifically by focusing on spatial geometry—you can get massive models' performance out of much smaller, more efficient architectures. Definitely worth a deep dive if you're into computer vision!