MoE architecture for video generation, with 3 billion parameters but only activating 300 million, this wave of efficiency optimization indeed hits the pain points of embodied intelligence.

Coin World News, Ant LingBot announced the open-source release of LingBot-Video, the world's first video generation foundation model for embodied intelligence. The model is based on a mixture-of-experts (MoE) architecture, designed to improve video generation efficiency. LingBot-Video has a total of 3 billion parameters, with only about 300 million parameters activated during generation. Compared to traditional dense architectures of similar parameter scale, inference efficiency is improved by approximately 3 times. This design enables the model to achieve visual expressiveness from large-scale parameters while being more suitable for the efficient inference requirements of embodied intelligence.