AntGroup's LingBot-Video Open Source: The World's First Video Foundation Model for Embodied Intelligence!

On July 9, Ant Group's LingBot-Video was open-sourced, which is the world's first open-source video generation foundation model based on the Mixture-of-Experts (MoE) architecture and tailored for embodied intelligence. The model redefines the video pre-training paradigm around the core needs of robots and embodied intelligence, achieving systematic improvements in reasoning efficiency, physical plausibility, action understanding, and task completion, providing a new open-source foundation for video foundation models to move from digital content creation to embodied intelligence.

On the benchmark RBench, jointly released by Peking University and ByteDance, LingBot-Video achieved a total score of 0.620, surpassing Wan2.6 (0.607), Seedance1.5Pro (0.584), and Cosmos3Super (0.581). As a comprehensive evaluation benchmark for robotic operation videos, RBench focuses on whether the model can generate robot behaviors that comply with real-world physical laws. This result indicates that LingBot-Video is better at maintaining the rationality of actions and the integrity of task execution when generating videos related to robots.

(Figure caption: LingBot-Video shows the best performance on RBench)

To further verify LingBot-Video's ability to model changes in the physical world, Ant Group evaluated it on an internal benchmark from two dimensions: general quality and embodied domain. The results show that compared to five open-source models, including NVIDIA Cosmos3, Wan2.2A14B, LongCat-Video, Hunyuan Video1.5, and LTX-2.3, LingBot-Video performs better than major baseline models in the embodied domain.

(Figure caption: Comprehensive evaluation shows that LingBot-Video demonstrates stronger physical understanding and action consistency in embodied scenarios)

In recent years, video generation models have made rapid progress in visual quality, smoothness, and creative expression. However, for embodied intelligence, a video that looks realistic and has smooth movements may not reflect real physical laws, making it difficult to support continuous prediction, planning, and task execution for robots. At the same time, embodied intelligence also requires models to have higher reasoning efficiency to adapt to real-time interaction and control loops.

This has led video generation to develop along two different directions: one heading toward cinemas, serving content creation; the other heading toward robots, serving the understanding, prediction, and interaction with the physical world. LingBot-Video represents Ant Group's important exploration into a new path for video generation tailored for embodied intelligence.

LingBot-Video has made systematic innovations in architecture, data, and training.

In terms of architecture, LingBot-Video adopts a DiT + MoE design, replacing traditional dense architectures with MoE. This allows the model to scale its capacity while controlling the cost of each inference. The 30B parameter model activates only about 3B parameters during generation, offering approximately three times the reasoning efficiency compared to a dense architecture of the same parameter size. This design enables the model to gain visual expressiveness from large-scale parameters while being more suitable for the efficiency requirements of embodied intelligence.

In terms of data, LingBot-Video built a data profiling engine, introducing robot-related data such as VLA, VLN, and Ego on top of massive internet videos, covering scenarios such as dexterous operations, robot mobility, and first-person interactions, with a total scale of 70,000 hours of embodied data. These data help the model learn the relationship between actions and environmental changes, rather than just learning the surface texture and visual style of videos.

In terms of training, LingBot-Video introduced a multi-dimensional reinforcement learning reward system. In addition to conventional metrics such as aesthetics, prompt following, and motion consistency, the model further aligns with physical plausibility and task completion, making the generated results more in line with real-world laws and closer to the needs of robots completing tasks in the real world.