Open-Source Robotic Video Foundation Model Released

dr_cintas · x · 2026-07-09

Robbyant has open-sourced LingBot-Video, the first MoE video foundation model designed specifically for robotics. The model has a total of 30B parameters but only activates 3B during inference. Officially, it is about 3x faster than dense models on long inputs, making it ideal for real-time robotic planning.

Built on a DiT + MoE architecture, the model was pre-trained on internet video data and further trained with 70,000 hours of embodied data, aiming to enhance its understanding of the physical world and robotic actions.

Related event: Ant Group Open-Sources LingBot-Video for Embodied AI(26 posts)→

Original post →

More from Embodied

Embodied channel →