Sand.ai open-sources MAGI-2, a 114B MoE video model
Beijing lab Sand.ai released Apache-2.0 weights for a 114-billion-parameter sparse video model that generates 1080p clips with synchronized audio.
Chinese video-generation startup Sand.ai published open weights and inference code on Wednesday for MAGI-2-preview, a 114-billion-parameter model that produces ten-second clips with a jointly generated soundtrack. The release is unusual for its architecture: MAGI-2 is a mixture-of-experts video model, activating roughly 6 billion parameters per token out of its 114 billion total.
What shipped
The model card describes a single-stream transformer built on what Sand.ai calls MagiMoE, with thousands of expert units per layer and multi-head expert routing — an attempt to import the sparse-scaling playbook that reshaped large language models into diffusion-style video generation, where nearly all frontier systems remain dense. Video and audio share the same attention stack rather than being stitched together by a separate voice or sound model, which the lab credits for lip-sync, sound-effect and ambience alignment.
Generation runs as a two-stage pipeline: a low-resolution preview pass followed by a refiner that upscales to 1088×1920. Only the fixed ten-second length is supported. The published checkpoints total roughly 300GB across the transformer stages, a Qwen3.5-27B text encoder, and video and audio VAEs, and the reference implementation expects eight NVIDIA Hopper GPUs. Everything is licensed under Apache 2.0, including the code, with a Docker image provided. Chinese outlet QbitAI put the inference cost of a ten-second 1080p clip at around half a dollar on an eight-H100 node, and noted the preview places sixth on Artificial Analysis' video-generation leaderboard despite its small active-parameter count.
Why it matters
Open video weights have been arriving steadily from Chinese labs, but almost all of them are dense models in the single-digit-to-low-tens-of-billions range, because video generation is bandwidth-bound and MoE routing is awkward to serve. Sand.ai's claim is that sparsity transfers: a model with a 114-billion-parameter knowledge capacity can be run at roughly the compute cost of a 6-billion-parameter one. If that holds up under community reproduction, it changes the cost curve for open video generation the way sparse LLMs did for text — and it lands days after MiniMax opened its H3 video weights, tightening a competition in which the highest-quality freely downloadable video model is now consistently Chinese.