NextFin News — Ant Group’s embodied intelligence subsidiary, Robbyant, open-sourced its LingBot-Video foundation model on Thursday, marking the industry's first Mixture-of-Experts architecture tailored for robotics video generation.
The model integrates over 70,000 hours of physical interactive data to redesign video pre-training paradigms for robotic task execution, precision physics, and operational reasoning. Built on a Diffusion Transformer combined with a Mixture-of-Experts (DiT+MoE) framework, the system maintains a 30-billion total parameter scale but activates only 3 billion parameters during active generation. This dynamic routing optimizes computing allocation, delivering execution speeds roughly three times faster than dense architectures of equivalent size. On the industry standard RBench evaluation platform, LingBot-Video secured a top score of 0.620, outperforming rival open-source architectures including Wan 2.6 and Cosmos 3.
The release marks a pivot from commercial digital content creation toward foundational physical world simulators. By providing open-source access to efficient spatial and action-prediction video layers, Robbyant targets an essential infrastructure layer required to lower downstream training friction for industrial automation developers.
Explore more exclusive insights at nextfin.ai.
