Ant Lingbo Open-Sources LingBot-Video, World's First MoE Video Foundation Model for Embodied AI

Ant Group's embodied AI subsidiary Ant Lingbo (Robbyant) has officially open-sourced LingBot-Video, the world's first large-scale Mixture-of-Experts (MoE) video foundation model designed specifically for embodied intelligence. The model is positioned as a video physics engine for robotics.

Unlike general-purpose video generation models that optimize for visual quality, duration, and cinematography, LingBot-Video prioritizes whether the generated video adheres to physical laws. The output must teach robots correct real-world physics — objects should not clip through each other, liquids should not float in mid-air, and movements must follow gravity and inertia.

On the architecture side, LingBot-Video adopts MoE with 30B total parameters but activates only approximately 3B per inference. This achieves 3.18x faster inference compared to a Dense 30B model at 1M token length, significantly reducing computational cost.

For training data, the model incorporates over 70,000 hours of embodiment-oriented footage covering robot manipulation, navigation, and first-person perspectives. The team employed a five-stage progressive curriculum, starting from low-resolution static images and gradually advancing to high-definition long-sequence video, with a multi-dimensional reward system incorporating physical plausibility and task completion.

In benchmarks, LingBot-Video achieved SOTA among open-source competitors on TI2V tasks, ranking first in both general quality and embodied domain scores. It also surpassed general-purpose video generation baselines on the RBench evaluation.

Ant Lingbo positions LingBot-Video as a Data Engine, Policy Evaluator, and Action Planner for the robotics community, aiming to provide a low-cost, repeatable physical world simulation environment for robot training.

The model is now open-source on GitHub, with the technical report on arXiv and model weights available on Hugging Face and ModelScope.