LingBot Video activates only ~3B of its 30B parameters per inference, trained on 70,000 hours of robot footage with a reward that flags physics violations, centering robot training over visual polish.
Ant Group's Lingbo unit open-sourced LingBot-Video on July 8 with a different success criterion than most video-generation releases. Rather than judging output by how real it looks, the lab trained the model with a reward that asks whether a robot learning from it would still get the physics right.
Where general video models optimize for aesthetic quality and prompt adherence, LingBot-Video trains with a multi-dimensional reward that explicitly flags physics violations: no interpenetration between objects, plausible persistence under occlusion, consistent inertia. A video clip of a robot grasping a cup is only useful as training data if the cup behaves like a cup when gravity is applied, the table top holds, and the hand does not pass through the handle. The lab treats this as the gap between video for viewers and video for robots, and the reward signal as the bridge.
The architecture is a Mixture-of-Experts design: 30 billion total parameters, with only about 3 billion activating per inference. The MoE design keeps inference cheap enough that the physics-aware training loop can run at video-token scale. The training corpus drew on more than 70,000 hours of embodied-relevant video, including robot manipulation, navigation, and egocentric scenes, a scale that is large for the embodied subfield but modest by general video model standards.
The release spans three primary channels: the GitHub repository robbyant/lingbot-video, the Hugging Face model card robbyant/lingbot-video-moe-30b-a3b, and the arXiv preprint 2607.07675, titled "Scaling Mixture-of-Experts Video Pretraining for Embodied Intelligence." Independent coverage from qbitai, chinaz, and tech.china.com confirms the timing and the headline numbers, but none of the three add independent benchmark data.
Ant Lingbo reports that LingBot-Video surpasses general video-generation baselines on RBench, an embodied-video benchmark. The claim is the lab's own evaluation, and the arXiv paper is a preprint rather than peer-reviewed work. The headline performance figures should be read as a positioning argument rather than settled fact until independent labs reproduce the result.
The unresolved question is whether embodied video pretraining closes the gap to downstream vision-language-action (VLA) policies, the mapping from camera input to robot motion, and to world-model training. Most robot foundation models still train on expensive teleoperation data, where a human physically guides a robot arm through a task and the trajectories are logged. A physics-aware video pretraining step could supply a cheaper source of structured visual priors, but only if generated trajectories transfer reliably to real robots. Ant Lingbo has not yet named a downstream robot platform that will train on LingBot-Video output, and the paper does not claim a finished policy stack.
General-purpose video models are evaluated by human viewers on aesthetic quality and prompt adherence. LingBot-Video is being evaluated on whether a robot that learns from the output would still succeed at the underlying task. The reorientation, more than the MoE architecture or the parameter count, is the part of the release that still has to land in real policy training.
LingBot-Video's open question is whether its physics-aware training objective transfers to real robot training data. Independent reproduction of the RBench result, or a published VLA policy built on the model, would settle the bet. Right now, the release stands as the bet itself: physics-aware video generation, distributed openly, in the hope that the embodied-AI field decides this is the kind of generated data it wants to learn from.