Sand.ai Open-Sources What It Calls the First 100B-Parameter MoE Video Generation Model

Sand.ai has announced the open-sourcing of its 100-billion-parameter Mixture-of-Experts (MoE) video generation model. According to QbitAI, it is the world's first 100B-scale MoE video generation model, pushing open-source video synthesis into the sparse-model era.

The model carries 114B total parameters, but its MoE architecture activates only 6B parameters per inference, cutting compute overhead substantially while aiming to preserve generation quality.

On capability, it generates 10-second 1080p videos, with a reported cost of about 0.5 yuan per generation — a price point that dramatically lowers the barrier to high-quality AI video.

Open-sourcing means developers can deploy, fine-tune, and reproduce the model in their own environments, giving individuals and small teams their first access to a 100B-scale video model that previously lived behind closed APIs.

The move signals that video generation is following the same trajectory as large language models: from closed APIs to open weights, and from dense models to sparse MoE designs that trade quality for much cheaper inference.

With only 6B active parameters, community adaptation becomes a realistic next step — quantization, fine-tuning, and inference optimization for consumer hardware are among the most anticipated efforts.

What to watch next: community benchmarks on temporal consistency, motion smoothness, and text adherence, and whether Sand.ai follows up with larger or domain-specific versions built on the same architecture.