LingBot-Video: A New Open-Source MoE Model for Embodied Video Generation

Robbyant has released LingBot-Video, an open-source Mixture-of-Experts (MoE) video generation model aimed at bridging video synthesis with embodied intelligence. The full release—technical report, code, model weights, and prompt rewriters—is available under an Apache 2.0 license. Training combined large-scale web video data with over 70,000 hours of embodied data, using a multi-reward setup for aesthetics, physical plausibility, and task completion.

The MoE design delivers roughly 3x faster inference than dense models. The suite includes a 1.3B dense model plus a 30B-A3B MoE model with a refiner, alongside two Qwen3.6-27B-based prompt rewriters. A three-stage inference workflow—prompt rewriting, automatic negative-prompt generation, and unified inference—supports both diffusers and SGLang Diffusion backends, with multi-GPU FSDP sharding for large checkpoints.

As of July 9, 2026, LingBot-Video tops the RBench leaderboard with an average score of 0.620, outperforming open-source peers like Cosmos3 Super, LongCat-Video, Wan 2.2, and HunyuanVideo 1.5, and staying competitive with closed-source models such as Wan 2.6, Seedance 1.5 pro, and Veo 3—particularly in manipulation, long-horizon, and quadruped tasks.

For developers building robotics and embodied AI systems, LingBot-Video offers an open, physically grounded video generation tool with flexible deployment options across hardware configurations.