Ant Group Open-Sources LingBot-World 2.0 AI World Models Including 1.3B Consumer GPU Variant

Ant Group's Lingbo Technology has open-sourced three new variants of its LingBot-World 2.0 real-time interactive world model, expanding on its previous 14B parameter release. The new additions include a lightweight LingBot-World 2.0 Small version with 1.3 billion parameters, alongside Bidirectional and Causal Pretrain model variants. By optimizing a 1.3B parameter interactive world model for single consumer-grade GPUs, the release makes real-time generative world simulation accessible to individual developers and small research teams. This significantly lowers the hardware barrier for exploring interactive AI environments, inference optimization, and real-time scene generation. LingBot-World 2.0 supports continuous generation over long durations while dynamically reacting to character actions like attacking or spellcasting, as well as weather events like rain or snow, achieving up to 720P at 60FPS on higher computing setups. Technically, the Bidirectional variant serves as a high-quality distillation source for training, while the Causal Pretrain variant provides a base for post-training, evaluation, and fine-tuning.