MiniMax released Music 3.0 on August 13, an open-weights music generation model that writes complete songs — not loops, not 30-second clips, but full arrangements up to five minutes long — from a lyric sheet and a text description of the sound you want. The model card is live on Hugging Face under MiniMaxAI/MiniMax-Music3, and the company published a companion technical writeup the same day.
A two-model system built for structure
Most AI song generators still struggle with the thing that makes a song feel like a song: it has to go somewhere. A verse should sound different from a chorus; a bridge should feel like a turn, not a repeat. Music 3.0 attacks that problem with a split architecture — an 8-billion-parameter global language model, initialized from Qwen3.5-8B, that tracks structure and lyrical content across the entire song, paired with a smaller 600-million-parameter local model that fills in moment-to-moment acoustic detail. A separate 2.4-billion-parameter flow-matching module and a 123-million-parameter Flow-VAE handle the actual audio synthesis, outputting 32kHz stereo.
The practical upshot is control most open music models don’t offer: explicit section tags like [Verse], [Chorus], and [Bridge] let a user dictate arrangement directly, and what MiniMax calls “Structured Captions” allow a different mood or instrumentation description per section rather than one blanket prompt for the whole track. You can ask for a track that opens sparse and builds into a full arrangement by the second chorus, and the model treats that as a structural instruction rather than an averaged-out compromise.
Open weights in a field that’s mostly closing up
The more consequential detail for working musicians and toolmakers might be the licensing, not the architecture. Suno, Udio, and ElevenLabs’ music platform are all API- or app-gated — you generate through their service, on their terms, with no access to the underlying model. MiniMax is publishing the weights alongside API access through platform.minimax.io, meaning anyone with the hardware can run Music 3.0 locally, fine-tune it, or build it into a pipeline without a subscription or a rate limit. That’s consistent with MiniMax’s pattern on its other recent releases (the H3 video model shipped the same way), and it’s part of why this site has been tracking MiniMax’s Hugging Face activity closely — a lab that opens its weights gets adopted into tool chains in ways an API-only competitor doesn’t.
What it’s actually useful for
Five minutes is long enough to cover most pop, folk, and electronic song structures without looping — a real gap in a field where a lot of “AI songs” are extended two-minute ideas. For creative technologists, the more interesting use may not be finished songs at all: the section-level control makes Music 3.0 a plausible backing-track generator for game and installation audio, where you need music that can shift mood on a cue without a human composer re-scoring every state change.