Ant Group has open-sourced LingBot-Video, described as the world’s first
open-source video foundation model for embodied intelligence built on a
Mixture-of-Experts (MoE) architecture. By replacing traditional dense
architectures with MoE to expand capacity while controlling per-inference cost,
the 30B-parameter model activates roughly 3B parameters during generation—about
three times the inference efficiency of a comparable dense model—aiming to
combine large-scale visual expressiveness with the efficient inference required
for embodied intelligence.
