LingBot-Video, an open-source video generation model for embodied intelligence, is an important tool for Ant Group.

Antminer Lingbo Technology has open-sourced LingBot-Video, the world's first video generation model based on the MoE architecture and geared towards embodied intelligence. The model has a total of 30 bytes of parameters, with only about 3 bytes activated during inference, making it three times more efficient than comparable Dense architectures. The team built a 70,000-hour embodied dataset and introduced a multi-dimensional reinforcement learning reward system to enhance physical plausibility and task completion.