LingBot-Vision may lower vision costs for AI agents
Ant Group released LingBot-Vision, an open-source . The weights are available under the in four sizes: 21M ViT-S, 86M ViT-B, 0.3B ViT-L, and 1.1B ViT-g. The model follows the DINO family of self-learning s and uses boundary-driven masking, where likely object-boundary tokens are hidden so the model cannot solve the task by copying nearby flat areas.
It uses no labels, no text supervision, and no external edge detector. On NYUv2 , the 1.1B model reached RMSE 0.296, better than DINOv3-7B at 0.309 and V-JEPA 2.1 2B at 0.307. The 0.3B ViT-L reached RMSE 0.310, almost the same as DINOv3-7B at 0.309, with about 23 times fewer parameters.
ViT-L is about 0.6GB in fp16, and the provided loader is aimed at feature extraction, , , and tracking, not chat. It does not win everywhere: ImageNet trails DINOv3 for the largest model and ViT-L.
Key points
- Ant Group released four LingBot-Vision s under Apache-2.0.
- The 0.3B ViT-L nearly matched DINOv3-7B on NYUv2 with about 23 times fewer parameters.
- The 1.1B ViT-g had the best NYUv2 RMSE in the listed comparison at 0.296.
- ViT-L is about 0.6GB in fp16, making it a possible lower-cost vision component.
- The release targets feature extraction, , , and tracking, not chat.