Ant Lingbo open-sources 1.3B lightweight LingBot-World 2.0 world model that runs in real time on a single GPU

Ant Lingbo has open-sourced a 1.3B-parameter lightweight LingBot-World 2.0 world model, following the 14B version released in July. It runs real-time world generation on a single consumer GPU. Unlike a pruned checkpoint, it was trained end-to-end with causal pretraining using MoBA attention masks, consistency distillation, and DMD distribution matching distillation.

The release includes the causal pretraining base and a bidirectional teacher for community post-training, distillation, and scene adaptation. AMD senior VP Jack Huynh demonstrated LingBot-World 2.0 during his IFA Berlin opening keynote.

Original sources (Chinese)