Ant Group Releases LingBot-World-Infinity 14B Causal World Model

Ant Group's robotics team Robbyant released LingBot-World-Infinity (LingBot-World 2.0), a 14B-parameter causal video generation model functioning as an interactive world simulator. Its core innovation is Mixture of Bidirectional Autoregressive attention masks (MoBA), combined with distribution matching distillation on long self-generated trajectories, addressing long-range drift issues like texture blur and geometric deformation common in interactive world models. A director-pilot agent framework is included: events proposed by VLMon are rendered by a diffusion Transformer. The paper demonstrates a continuous 60-minute session covering 20 scenes. The release is limited: only one checkpoint, 480P reference scripts, no deployment code or quantitative benchmarks, under CC BY-NC-SA 4.0. For teams studying causal world models and embodied AI, this provides a new open-source baseline, but its limited release means replication and further development require engineering effort.