What is lingbot-video?

LingBot-Video is an open source video generation model built for what the project calls embodied intelligence, meaning AI systems meant to understand and act in the physical world, like robots. The idea is that if a model can generate realistic video of objects and actions, it can also learn something useful about how the physical world behaves, not just how to make pretty clips. The model uses a design called Mixture of Experts, where instead of running one giant network for every request, the system routes each piece of work to a smaller subset of specialized parts. The README says this approach was built from scratch and gives roughly three times faster inference than a similarly capable dense model. The model was trained on a large mix of general web video plus more than 70,000 hours of video specifically showing embodied, real world tasks, and it was tuned with a reward system that scores outputs on visual quality, physical plausibility, and whether the requested task actually looks completed. Two main model sizes are offered: a smaller 1.3 billion parameter dense model, and a larger 30 billion parameter Mixture of Experts model paired with a refiner stage for higher quality output. Both support generating images from text, videos from text, and videos from a starting image plus text. There are also separate helper models that rewrite a user's casual prompt into the structured JSON format the main model expects, since it does not work well with plain natural language directly. To run it, a user clones the repository, sets up a Python virtual environment, and installs the listed dependencies, including PyTorch and the Hugging Face libraries the project depends on. Generating a video means first running the prompt rewriter to turn a plain description into JSON, optionally pruning a negative prompt automatically, and then running the main inference script with either a direct backend or a faster SGLang based backend. Scripts are included for both single GPU and multi GPU setups, with the multi GPU option splitting the model in memory to handle its larger size.