Skip to main content
Back

SandAI's MAGI-2 Generates 1080p Video With Sound at One-Tenth the Cost

SandAI released MAGI-2 Preview, a 114B open-source mixture-of-experts video model that activates only 6B parameters per token and generates 10-second 1080p clips with synchronized audio.

  • SandAI open-sourced MAGI-2 Preview, a 114B MoE video model activating only 6B parameters per token
  • Generates 10-second 1080p clips with synchronized audio from text or text plus a still image
  • Ranks #6 on the Artificial Analysis video leaderboard at claimed 1/10 the cost of mainstream models
  • Uses Ultra-Fine-Grained MoE: 12 heads each pick 6 of 256 experts, activating 72 per token
  • Requires 8 NVIDIA Hopper GPUs and a 307 GB checkpoint download from Hugging Face
  • Apache 2.0 license; distilled fast version coming, per the tech report

SandAI has released one of the largest open-source video generation models to date. MAGI-2 Preview is a 114B-parameter unified audio-video generation model that activates just 6B parameters per token, producing 10-second clips at 1080p with the soundtrack generated jointly rather than dubbed on afterwards. Weights, inference code, and a technical report are all public under Apache 2.0.

The team claims roughly one tenth the generation cost of mainstream models at competitive quality, with the model landing at #6 on the Artificial Analysis video generation leaderboard despite its small activated footprint. A 100B+ open model with genuinely sparse compute is what makes this release worth digging into.

What actually shipped

MAGI-2-preview generates 10-second clips with synchronized audio from either a text prompt or a text prompt plus a still image, using a sparse mixture-of-experts design with 114 billion total parameters but activating only 6 billion per token. Generation runs in two stages: magi2_preview denoises the clip at low resolution, and magi2_refiner takes that result up to 1080p.

The preview stage generates 512x896 and the refiner takes that to 1088x1920. The VAE stride forces every generated dimension to be a multiple of 16, so 1080p is a delivery target that gets rescaled from the 1088-wide native shape.

What you get out of the box:

  • Text-to-video (T2V) and image-to-video (I2V) in a single pipeline
  • Joint audio generation muxed into the output MP4 via ffmpeg
  • Two transformers: a preview stage (228 GB) and a refiner stage (14 GB)
  • A Qwen3.5-27B text encoder, plus video and audio VAEs
  • A distilled turbo VAE decoder for fast latent decoding
  • Apache 2.0 license on both code and weights

The full checkpoint bundle weighs in at roughly 307 GB. Fixed clip length is a real constraint, and 10 seconds is the only duration the model currently supports.

Shredding tokens across 256 experts

The headline architectural idea is what SandAI calls Ultra-Fine-Grained MoE. Instead of routing whole tokens to a small pool of experts (the standard LLM setup), MAGI-2 shreds the hidden state and routes each piece independently. A 3,072-dimensional hidden state is split into 12 heads. Each head picks 6 of 256 experts, so a single token activates 72 experts overall.