Robbyant, the embodied AI division of Ant Group, has officially open-sourced LingBot-Video, describing it as the world’s first large-scale Mixture-of-Experts (MoE) video foundation model designed specifically for embodied intelligence. Released under the Apache 2.0 license, the project aims to bridge the gap between traditional AI video generation and real-world robotics applications.
If you’re thinking about purchasing a new GPU, we’d greatly appreciate it if you used our Amazon Associate links. The price you pay will be exactly the same, but Amazon provides us with a small commission for each purchase. It’s a simple way to support our site and helps us keep creating useful content for you. Recommended GPUs: RTX 5090, RTX 5080, and RTX 5070. #ad
Unlike conventional text-to-video models that prioritize visual aesthetics, LingBot-Video was built from the ground up to understand physical interactions, object dynamics, and task-oriented actions. The model is intended to provide a stronger foundation for robotics, simulation, and embodied AI while remaining capable of high-quality video generation.
A Mixture-of-Experts Architecture
One of LingBot-Video’s biggest innovations is its sparse Mixture-of-Experts (MoE) architecture. The model contains approximately 30 billion parameters, but only about 3 billion parameters are activated during inference, allowing it to scale its capacity without requiring the full computational cost of a dense model.
According to the research team, this design improves inference efficiency while maintaining strong video generation capabilities.
Designed for Physical Intelligence
Rather than training exclusively on internet videos, the researchers supplemented the training data with extensive robotics-oriented footage, including:
- Robot manipulation
- Navigation tasks
- Egocentric viewpoints
- Action-focused videos
The project also introduces a multi-dimensional reward system that evaluates not only image quality and prompt adherence but also physical plausibility and successful task completion. These additions are intended to help the model better understand how objects and environments behave in the real world.
Open Source Under Apache 2.0
LingBot-Video has been released as an open-source project under the Apache 2.0 license, making it available for both research and commercial development.
The GitHub repository includes:
- Source code
- Model weights
- Documentation
- Technical report
- Installation instructions
This release allows researchers and developers to experiment with the model and build new applications on top of it.
Part of a Larger LingBot Ecosystem
The release of LingBot-Video coincides with Robbyant’s broader push into embodied AI. Alongside the video model, the company recently introduced several new open-source projects, including LingBot-World 2.0, LingBot-VA 2.0, LingBot-VLA 2.0, and LingBot-Depth 2.0, all targeting different aspects of robotics, world modeling, and autonomous agents.
Final Thoughts
While most AI video models today focus on content creation, LingBot-Video represents a different direction. By emphasizing physical reasoning, action understanding, and efficient sparse MoE inference, the project demonstrates how video foundation models can evolve beyond entertainment and become building blocks for embodied AI systems.
Although the model’s primary target is robotics research, its open-source release may also attract the attention of AI creators interested in experimenting with next-generation video generation technology.
Resources
- https://technology.robbyant.com/lingbot-video
- https://huggingface.co/robbyant/lingbot-video-moe-30b-a3b
- https://github.com/robbyant/lingbot-video
Leave a Reply