MiniMax H3

MiniMax H3 gives teams several ways to specify a short shot: first and last frames, or a package of image, video and audio references. It generates video and stereo sound together, making it worth considering for permitted cutscenes and promotional footage. Released weights allow local integration at 768p; the official context-processing and 2K stages remain hosted. The catch for a worldwide game is the community license: its territorial restrictions extend to generated outputs. Hosted processing can also block pornography. Settle the license and content-policy fit, then compare a small batch of acceptable shots before spending on a larger render pipeline.

Commercial use & license

The MiniMax H3 Community License excludes the EU, UK, US and South Korea; it also bars use of outputs there. Revenue above US$20 million a year from commercial products or services requires prior written authorization. Commercial UI attribution and the AUP apply. Confirm the hosted agreement separately; globally advertised API access does not resolve this community license’s output restrictions.

Adult content rules

Hosted context processing may block pornography. Local use remains bound by the community AUP, including consent, public AI disclosure and territorial rules, plus restrictions tied to regional social and moral standards. Local weights are not blanket permission for adult content.

Read the policy ↗

What you need

The local FL2VA and Ref2VA checkpoints generate four- to fifteen-second clips at a 768-pixel short side and 24 fps. Install the matching processor, tokenizer, encoder, visual VAE and audio VAE. Hardware needs depend on the configuration; official context processing and 2K regeneration require hosted services.

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 →