Robbyant open-sources LingBot-Video, an MoE model for robotics

Robbyant dropped a LingBot-Video collection on Hugging Face aimed at video generation for embodied AI rather than social feeds. The model card calls it 'the first open-source large-scale MoE (Mixture-of-Experts) video generation model dedicated to embodied intelligence,' designed to bridge video synthesis with physical world understanding.

The headline release is a 30B-A3B mixture-of-experts model, meaning 30 billion total parameters with 3 billion active per token, alongside a smaller 1.3B dense variant and a LoRA prompt rewriter. According to the card, it was trained on web video integrated with more than 70,000 hours of embodied data, and the reward signal during training weighted aesthetics, physical rationality, and task completion together rather than aesthetics alone. Everything ships under Apache 2.0 with a diffusers-compatible pipeline that supports text-to-image, text-to-video, and text-image-to-video.

The self-reported benchmark story is where it gets specific. On the RBench leaderboard the team cites, LingBot-Video posts an overall average of 0.620 as of July 9, 2026, and the domain breakdown is uneven in a telling way. Quadruped hits 0.758, humanoid 0.689, dual arm 0.639, and long-horizon tasks 0.634. Reasoning lags at 0.505 and multi-entity scenes are the weakest at 0.444.

The honest caveat is that these numbers come from the model's own page, not an independent leaderboard operator, and a strong RBench score for generated video is not the same as usefulness inside a live robot policy loop. What the release does not describe is the provenance of the 70,000-hour embodied dataset, and total training compute is not disclosed either.

For robotics teams the interesting bit is what the '~3x faster inference' claim, combined with a 3B active-parameter budget and an Apache 2.0 license, actually enables. Sim-to-real pipelines and quadruped or single-arm inspection stacks get an open building block that fits on a modest GPU cluster, and that is the kind of ingredient this corner of the space has been short on.