What is MAGI-2 Preview?
MAGI-2 Preview was published on 5 August 2026 by Sand AI and supersedes its MAGI-1. It is a mixture-of-experts transformer with about 114 billion parameters, roughly 6 billion active per token, that generates video and its soundtrack together from text or from text plus a still image. Sand AI calls it an intermediate research release.
Clips are 10 seconds long, the only duration supported. A preview stage generates at 512x896 and a 14GB refiner takes the result to 1088x1920. The full checkpoint is about 307GB, including a 56GB Qwen3.5-27B text encoder and the Wan2.2 video VAE, and inference needs eight NVIDIA Hopper GPUs. Code and weights are under Apache 2.0.
Is MAGI-2 Preview worth it?
MAGI-2 Preview gives the open community an audio-video model under a genuinely permissive licence, at a hardware price few can pay. On Artificial Analysis's text-to-video board with audio it ranks 16th of 32 at 1040 Elo, ahead of Kling 3.0 1080p (Pro) at the board's fixed 1000 and LTX-2.5 Fast at 942, but 120 points behind MiniMax H3 at 1160.
The licence is the reason to care. MiniMax H3, the only open model above it, requires a separate application for use in the US, the EU, the UK and South Korea. MAGI-2 is Apache 2.0, with no revenue cap and no excluded regions.
The hardware is the obstacle. The code expects eight Hopper GPUs, and even then the preview and refiner stages are moved in and out of GPU memory because, at 1080p, neither fits on an 80GB card beside the other's working data. The text encoder alone is a 27-billion-parameter language model.
Sand AI is candid that this is a preview. Clips are fixed at 10 seconds, prompts work best as long structured JSON captions from a separate language model, no generation times are published, and the blog describes its evidence for consistent characters and matching subtitles as informal observation rather than controlled tests.
Where it earns its place
A university video-generation lab with an 8x H100 node
Study and fine-tune a large mixture-of-experts video model without licence negotiations
Pull the Docker image and the 307GB checkpoint; Apache 2.0 permits publishing derivatives, and deterministic mode makes runs bit-exact at some cost in speed.
A Berlin advertising studio that cannot use territorially licensed models
Generate 10-second spots with synchronised sound on its own hardware
Rent an eight-GPU Hopper node, turn briefs into JSON captions with an instruction-following LLM using the shipped system prompts, and render at 1088x1920. Unlike MiniMax H3, no EU licence application is needed.
Before you commit
- Setup
- docker pull sandai/magi-2-preview, download the 307GB checkpoint with hf download and run scripts/run_demo.sh under torchrun on eight GPUs; ffmpeg is needed to add the audio track.
- Integrations
- Native PyTorch pipeline with MagiAttention and MagiCompiler and a Docker image per commit; no Diffusers or ComfyUI support is listed.
- Data and privacy
- Self-hosted only for now, so prompts and footage stay on your hardware; training data size is not published.
- Cost at scale
- No API price yet. Self-hosting means an eight-GPU Hopper node per worker, and Sand AI publishes no clips-per-hour figure.
- Limits
- 10-second clips only; dimensions are multiples of 16; text-to-video and image-to-video only; a research preview with informal quality claims.
What’s good
- Apache 2.0 weights with no revenue or regional limits
- Video and audio generated together
- 16th of 32 on Artificial Analysis's audio-video board
- Refiner outputs 1088x1920
What’s not
- Needs eight NVIDIA Hopper GPUs
- About 307GB of weights to download
- Fixed 10-second clips
- No hosted API or published speed figures yet
Best for
Research labs and well-funded studios with an eight-GPU Hopper node who want a permissively licensed audio-video model.
MAGI-2 Preview alternatives
MAGI-2 Preview: common questions
- Is MAGI-2 Preview worth it?
- MAGI-2 Preview is the strongest video-with-sound model you can use under Apache 2.0, and the first thing it asks for is eight Hopper GPUs. We rate it 3.7 out of 5.
- How much does MAGI-2 Preview cost?
- Free weights under Apache 2.0. No hosted price yet; Artificial Analysis lists the API as coming soon.
- What are the downsides of MAGI-2 Preview?
- Needs eight NVIDIA Hopper GPUs; about 307GB of weights to download; fixed 10-second clips; no hosted API or published speed figures yet.
- Who is MAGI-2 Preview best for?
- Research labs and well-funded studios with an eight-GPU Hopper node who want a permissively licensed audio-video model.
- How much does MAGI-2 Preview cost for a team or in production?
- No API price yet. Self-hosting means an eight-GPU Hopper node per worker, and Sand AI publishes no clips-per-hour figure.
- Is MAGI-2 Preview safe for company data?
- Self-hosted only for now, so prompts and footage stay on your hardware; training data size is not published.
- What are the best MAGI-2 Preview alternatives?
- The closest alternatives we have reviewed are MiniMax H3 (Hailuo 3.0), LTX-2.5, Kandinsky 6.0 Video and Wan 3.0.
Answers checked Nov 7, 2026.