The world's first video prototype for embodied intelligence, Ant Lingbo LingBot-Video open source

IT House reported on July 9 that today, Ant Lingbo Technology officially open sourced LingBot-Video. This is the world's first open source video generation basic model based on the Mixture-of-Experts (MoE) architecture and oriented to embodied intelligence. IT Home attaches an official detailed introduction as follows: In the past few years, video generation models have made rapid progress in image quality, fluency and creative expression. But for embodied intelligence, a video that looks realistic and moves smoothly may not reflect the real physical laws, and it is often difficult to support the continuous prediction, planning and execution of tasks by robots. At the same time, embodied intelligence also requires models to have higher reasoning efficiency to adapt to real-time interaction and control closed loops. This has also led to two different directions of evolution in video generation: one leads to theaters, serving content creation; the other leads to robots, serving the understanding, prediction, and interaction of the physical world. LingBot-Video is our important exploration to open up a new route of video generation for embodied intelligence - re-designing the video pre-training paradigm around the core requirements of embodied intelligence to achieve systematic improvements in reasoning efficiency, physical rationality, action understanding and task completion. System innovation in architecture, data, and training For embodied intelligence, LingBot-Video has made system innovation in three aspects: architecture, data, and training: Architecture: DiT + MoE design We replace the traditional Dense architecture with MoE to expand model capacity while controlling the cost of single inference. LingBot-Video's 30B total parameter model only activates about 3B parameters when generating, which is about 3 times more inferential than the Dense architecture of the same parameter scale. This design allows the model to not only obtain the visual expression ability brought by large-scale parameters, but also be more suitable for the requirements of embodied intelligence for efficient reasoning. Data: Data profiling engine and 70,000 hours of embodied data We built a data profiling engine, based on massive Internet videos, and further introduced VLA, VLN, and Ego