by Sand AI
Open-weights 114B MoE model from Sand AI that generates synchronized video and audio from text or image prompts.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
MAGI-2 Preview is a 114B-parameter unified audio-video generation model from Sand AI that activates about 6B parameters per token. It is built on a mixture-of-experts architecture called MagiMoE. The model card says the architecture, training system and data pipeline are described in a Sand AI tech blog post on scaling video generation efficiently.
It generates video from a text prompt (T2V) or from a prompt plus a still image (I2V), with sound generated alongside the video and muxed into the output. YouTube coverage describes clips of about 10 seconds, with a refiner stage reaching up to 1080p.
Weights are published on Hugging Face under Apache-2.0, with inference code in a separate repository. A creator review puts the checkpoint at roughly 228 GB and says it needs eight Hopper-class GPUs, which makes local use impractical for most individuals. It is labelled a preview release.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Open-source 114B MoE video generator with native audio, 10-second generations and refiner up to 1080p. 228 GB model needing eight Hopper GPUs, so impractical for consumers.
Closest neighbours by meaning, not just by category.
Model · ByteDance
ByteDance's open-weights video generation and editing model pairing a Qwen2.5-VL planner with a Wan2.2 renderer.
Model · Lightricks
Open-weights Lightricks video model with native audio, multi-shot generation, auto duration and up to 4K output.
Model · ByteDance
ByteDance Seed's multimodal video generation model, producing video from text and reference inputs.
Model · SenseTime
Open-weights 27B model from SenseTime that creates short videos by generating and rendering executable code.
Model · Hao AI Lab
Distilled MiniMax H3 reference-to-video-and-audio model from Hao AI Lab, generating in 8 steps instead of 50.
App
Open-source local web app for running video, image, audio, music and 3D generative models on low-VRAM GPUs.