Ant Group open-sources LingBot-Video for robot video generation

Ant Group's Lingbao Technology has open-sourced LingBot-Video, the first open-source video generation model for embodied intelligence based on MoE architecture. The model has 30B total parameters but activates only 3B per inference, making it 3x more efficient than dense models of similar size. It uses 70,000 hours of robot data and a multi-dimensional reinforcement learning system to improve physical reasoning and task completion. On the RBench benchmark, it scored 0.620, beating Wan2.6 and Seedance 1.5 Pro. The model is now available for robot action prediction and simulation.