LingBot-Video: Embodied Video Foundation Model

aigclink · x · 2026-07-10

LingBot-Video is described as the first open-source MoE video foundation model designed for embodied AI. Instead of focusing on creative content quality, it reimagines video pre-training specifically for robotics, detailing how its architecture, data, and training are tailored for physical interaction and long-horizon modeling.

The model features 30B total parameters with only 3B activated per token. It incorporates 70,000 hours of embodied data and physics-based reward designs. It reportedly leads the public RBench leaderboard and can serve as a foundational component for action-conditioned video generation, data generation, and policy evaluation.

Related event: Ant Group Open-Sources LingBot-Video for Embodied AI(26 posts)→

Original post →

More from Embodied

Embodied channel →