LingBot-Video introduces the first open-source, large-scale Mixture-of-Experts (MoE) video generation model specifically designed for embodied intelligence. It aims to bridge the gap between video synthesis and physical world understanding, offering a powerful tool for researchers and developers in AI. The project provides efficient MoE architectures, a robust data engine trained on extensive embodied data, and a multi-reward system that prioritizes aesthetics, physical rationality, and task completion, leading to state-of-the-art performance.

How It Works

LingBot-Video employs an efficient MoE architecture, scaled from scratch to balance computational capacity and cost, achieving approximately 3x faster inference compared to alternatives. Its training leverages a massive dataset comprising web videos and over 70,000 hours of embodied data, curated by a sophisticated data engine. The model is trained using a multi-reward system that optimizes for high aesthetic quality, physical plausibility, and successful task completion in generated videos.

Quick Start & Requirements

Installation involves cloning the repository, setting up a Python virtual environment, and installing dependencies via pip install -r requirements.txt and pip install -e .. Key prerequisites include Python >= 3.10 and a recommended PyTorch build (2.12.0.dev20260220+cu130) with corresponding torchvision. Other dependencies include transformers, diffusers, peft, json_repair, decord, and safetensors. CUDA is implicitly required for GPU acceleration. Official documentation and project pages are available for further details.

Highlighted Details

• Features an efficient MoE architecture offering ~3x faster inference.
• Trained on a massive dataset including 70,000+ hours of embodied data.
• Utilizes a multi-reward system for aesthetics, physical rationality, and task completion.
• Offers both Dense (1.3B) and MoE (30B-A3B) model variants, alongside prompt rewriters.
• Achieved top ranking on the RBench Leaderboard as of July 9, 2026.

Maintenance & Community

The project released its technical report, code, and models on July 9, 2026. Community interaction is facilitated through a WeChat group.

Licensing & Compatibility

This project is licensed under the Apache 2.0 License, which generally permits commercial use and integration into closed-source projects.

Limitations & Caveats

Inference requires structured JSON captions, necessitating the use of a bundled prompt rewriter. Multi-GPU inference with large MoE checkpoints demands substantial system RAM. The recommended PyTorch build is a development version, which may require specific environment configurations.