Researchers Open-Source LingBot-Video MoE Model for Embodied Robot Intelligence
A Hugging Face release and X posts are drawing attention to a 30B-parameter robotics video model that says it uses only 3B parameters at inference.
13121199626.7K
An open-source foundation model for robots. Show me something more hardcore than this, and I'll deactivate my account. Bots are coming!
SvpinoTECH#1893via XStory Brief
A Hugging Face collection page for LingBot-Video and accompanying posts on X from @_akhaliq present the project as an open-source Mixture-of-Experts video foundation model for embodied intelligence. The main claim in those posts is a 30B-parameter model with only 3B parameters active at inference, built on large-scale internet video pretraining and augmented with 70,000 hours of embodied data for robotics-style tasks.
Impressive work. 10:1 param-to-active ratio is well-suited for real-time robotics. Curious how it generalizes beyond RBench though, embodied environments are unforgiving.
SvenUrbanSciTECHvia X