Mini toolsLingBot Video MoE
By Robbyant
Model family: LingBot
LingBot-Video-MoE is a 30B parameter (3B active) Mixture-of-Experts video generation model from Robbyant, designed to bridge video synthesis with physical world understanding for embodied intelligence. It supports text-to-image, text-to-video, and text+image-to-video generation, along with a dedicated refiner model for higher quality output. The model consumes structured JSON captions produced by a companion prompt rewriter rather than raw prompts, with an auto-negative-prompt step to prune negatives per caption. Compared to dense architectures, the MoE design gives roughly 3x faster inference. LingBot-Video is trained on large-scale web video combined with over 70,000 hours of embodied data and optimized with a multi-reward system for aesthetics, physical rationality, and task completion. Released under Apache 2.0.
New
Video
Overview
LingBot-Video-MoE is an open source Mixture-of-Experts video generation model with 30B total and 3B active parameters, built for embodied intelligence. It supports text-to-image, text-to-video, and image-to-video generation, plus a refiner stage for higher quality output, and is trained on web video combined with over 70,000 hours of embodied data.
Last updated: July 8, 2026
