MAGI-2 Preview

MAGI-2 Preview generates ten-second clips from text or a still image, with audio produced alongside the video. Its downloadable weights and two-stage pipeline give teams control over local inference, including a refiner for 1080p output. That control comes with a serious hardware bill: the reference setup uses eight Hopper GPUs and roughly 307 GB of checkpoints. It suits research teams with that infrastructure; a single-GPU visual novel team has little reason to start here. The base release is available, while the promised distilled version remains pending. Before planning adult output, check the supporting components: Stable Audio Open carries restrictions on sexually explicit content.

Commercial use & license

The repository is Apache 2.0; supporting components retain their own terms. Stable Audio Open's community terms require commercial registration and a separate license above US$1 million annual revenue. Its AUP also applies. Check the full dependency set before commercial use.

Adult content rules

The documented audio pipeline includes Stable Audio Open. Its community license incorporates Stability's AUP, which prohibits sexual acts and other sexually explicit content even for self-hosted use. The MAGI repository's Apache label does not establish different rights for that component.

Read the policy ↗

What you need

Official reference setup: eight NVIDIA Hopper GPUs, Python 3.12, a recent CUDA toolkit, and ffmpeg to include the audio track. The model has 114B total parameters, 6B active per token, and supports ten-second clips. Allow roughly 307 GB for checkpoints and install the supporting-model dependencies.

Ranks are recommendations for the task described, based on reviewed documentation. Model ranks cover non-explicit development work. We have not run practical benchmarks. Check content policies and commercial terms separately. How the ranks work →