Ant Group launches LingBot-Video and LingBot-Vision for robotics

Ant Group has open‑sourced two foundational AI models aimed at embodied robotics. LingBot‑Video, released on July 9, is the first video‑generation model built on a Mixture‑of‑Experts architecture for robot‑related content. With a 30 billion‑parameter backbone that activates about 3 billion parameters during inference, it achieves higher efficiency and better physical plausibility than prior models, scoring 0.620 on the RBench benchmark and outperforming alternatives such as Wan2.6 and Cosmos3 Super. The model was trained on roughly 70,000 hours of robot‑centric video data and incorporates multi‑dimensional reinforcement learning rewards to improve task completion and physical realism.

The companion model, LingBot‑Vision, is a self‑supervised vision foundation model designed to enhance dense spatial perception for robotics. Its flagship ViT‑g/16 version contains about 1.1 billion parameters and uses a novel masked boundary‑modeling pre‑training strategy on a curated dataset of 161 million images. It delivers strong results on depth estimation tasks, recording an RMSE of 0.296 on the NYU‑Depth v2 benchmark and competitive performance on the KITTI benchmark, surpassing larger models such as DINOv3. Both models are released under an Apache 2.0 licence, providing the robotics community with open‑source tools for action prediction, simulation data generation, and world‑model research.