LingBot-Video是专为具身AI设计的Diffusion Transformer视频基础模型，采用MoE架构，为首个开源具身视频生成模型。旗舰30B参数每次仅激活3B，支持1M token序列，同参数量下推理效率约为Dense架构的3.18倍。模型通过因果关系、物体恒存、非穿透及材质真实等物理合理性评分来优化物理真实性，而非仅审美奖励。预训练数据包含70,000+小时具身视频（机器人操作、导航、第一人称动作）。在RBench上平均得分0.620，领先所有列出的开源和闭源模型，但部分指标仍不及Wan 2.6。
A first-of-its-kind open-source video generation model for embodied scenarios, using a Mixture-of-Experts design.
LingBot-Video from @robbyant_brain, a Diffusion Transformer video foundation model built for embodied AI.
- The flagship 30B model activates only 3B parameters per generation, at 1M-token sequences
- At the same total parameter scale it reaches ~3X the inference efficiency of a Dense architecture at 1M-token sequences (3.18× vs. Dense 30B)
- Aesthetic-only reward models are why video generators fake physics, and this paper basically says so. Physical plausibility is not treated as vibes here; it gets scored through causality, object permanence, non-penetration and material realism.
- The model was not trained only on regular internet videos. Its pretraining data also includes 70,000+ hours of embodied footage, covering robot manipulation, navigation, and first-person action videos, so the model sees examples where movement actually changes the world.
- On RBench, it reports a 0.620 average, ahead of every listed open and closed model, while still losing some columns to Wan 2.6.
🧵 1.
来源：Rohan Paul · x.com