LingBot-Video puts a 30B MoE behind embodied AI video

A new text-to-video model called LingBot-Video 30B-A3B has landed on Hugging Face, aimed squarely at embodied intelligence rather than general-purpose clip generation. The release pairs a diffusion transformer (DiT) backbone with a mixture-of-experts design, and it arrives under a permissive Apache 2.0 license (model card).

The naming tells the architectural story. Of its roughly 30 billion total parameters, only about 3 billion are active on any given forward pass — the "A3B" suffix. That sparse-activation approach is meant to give the model the capacity of a large network while keeping inference costs closer to those of a much smaller one.

Why it matters

Most open video models chase cinematic prompts and short social clips. Framing this one as a pretraining model for embodied intelligence points at a different use case: generating and predicting visual sequences that robots and agents can learn from, where a world model that anticipates how a scene unfolds is more useful than a polished trailer.

A few things stand out about this release:

• Sparse compute: 30B total parameters with only 3B active, via MoE routing.
• DiT foundation: a diffusion transformer backbone, now the dominant design for high-quality video generation.
• Open license: Apache 2.0 permits commercial use and fine-tuning without downstream restrictions.

Several practical details — resolution, frame rate, and maximum clip duration — aren't specified in the release record, so teams evaluating LingBot-Video for robotics or simulation work will want to test those limits directly. As a version 1.0 initial release, it establishes a baseline that later checkpoints can build on.