MiniMax Music 3
Five-minute songs, open weights — and from the H3 lab, a license with every clause except the territory wall
What it is
MiniMax's open-weights music generator: hand it lyrics with section tags and a music description, get back a complete, structured song — up to five minutes of 32 kHz 16-bit stereo WAV. The stack is a hybrid: an 8B global LLM for long-range musical structure, a 0.6B local LLM for frame-level acoustic detail, a 2.4B flow-matching module, a 123M Flow-VAE decoder, and an eight-layer RVQ tokenizer. Served officially through SGLang-Omni, with a Diffusers pipeline in review and ComfyUI tutorials already documented.
Why it's interesting
Our audio shelf needed this entry, and the watch list flagged it as 'reported Apache-compatible — needs a proper read.' The proper read says otherwise: the LICENSE is a bespoke MiniMax-Music3 Community License, not Apache — but it's also a markedly friendlier document than sibling MiniMax H3's. No territory clause: the grant that H3 denies to the US, EU, UK, and Korea, Music 3 extends worldwide. What remains is the familiar fine print — a binding AUP, mandatory 'MiniMax-Music3' UI branding on commercial products, and a $20M yearly revenue trigger for written authorization. Same lab, same month, two very different licenses; that contrast is worth an entry on its own.
Use cases
- Self-hosted full-song generation — lyrics in, mixed stereo master out, no per-generation API bill
- Music production workflows via ComfyUI and SGLang-Omni, on your own hardware
- A license case study: what the H3 lab writes when it isn't fencing off jurisdictions
Who it's for
Musicians and toolmakers self-hosting generation; commercial deployers, read the branding clause first
Setup
Moderate. ~11B parameters across four components — a single serious GPU serves it; official SGLang-Omni cookbook, ComfyUI workflows, and community GGUF ports available
Limitations & cautions
The AUP travels with every redistribution and the UI branding applies to any commercial product regardless of size; training data is undisclosed, which in music generation means the copyright provenance question is entirely unanswered; and output caps at five minutes / 9,000 acoustic frames per generation.
Editorial takeaway
The most interesting thing about this license is who wrote it: the same lab that walled four jurisdictions out of H3 shipped its music model to the whole world. Fine print is a choice, made release by release — which means it can be unmade.