Ant Group's Lingbo Technology has released three new open-source models under the LingBot-World 2.0 series. This follows the initial release of the 14B version in July. The new additions include the 1.3B parameter Small model, along with Bidirectional and Causal Pretrain variants.
The LingBot-World 2.0 system functions as a real-time interactive world model. It supports continuous generation for hours and reacts to environmental changes like rain or snow, as well as character actions such as shooting or spellcasting. With sufficient computing power, the model can achieve 720P resolution at 60 frames per second.
The 1.3B Small model is specifically optimized for consumer-grade single-GPU setups, allowing for real-time generation on local hardware. This design aims to lower the barrier to entry for individual developers, university laboratories, and small teams interested in exploring inference optimization and scenario applications.
The remaining two models serve specific development needs. The Bidirectional model is designed for training real-time world models by providing high-quality distillation sources. Meanwhile, the Causal Pretrain model acts as a foundation for post-training, model evaluation, and scene adaptation tasks.
