What the model does

MiniMax has released MiniMax-Music3, a text-to-music model, with open weights. The model takes two separate inputs: lyrics carrying section tags, and a detailed music description.

The output is a complete song of up to five minutes in a single generation, as 32 kHz, 16-bit stereo WAV. That distinction matters: many music models produce a few seconds at a time and stitch the pieces together, whereas this one returns the whole song at once, so section transitions do not read as seams.

The architecture

The structure has two layers. The first is a language model layer called Hybrid-LM, pairing an 8B global model with a 0.6B local model. The second is a continuous synthesis stack built on flow matching and a Flow-VAE.

That separation means song structure and sound texture are handled in different layers: the language model decides what comes when, while the synthesis layer produces the actual waveform.

Deployability

What shows this is not a research preview is what shipped on day one:

  • usable weights,
  • inference code,
  • three documented serving paths.

The model can therefore be set up and run today. Solo creators, indie studios and mid-market teams can ship on it directly. The intended areas include game development, advertising agencies, short-form video and creator tools, e-learning and podcasting.

The licence details matter

The MiniMax-Music3 Community License permits commercial use, but under three conditions.

First, the name MiniMax-Music3 must be displayed prominently in the product UI. Second, any organisation whose aggregate yearly revenue from those products exceeds $20 million must obtain separate prior written authorisation from MiniMax. Third, anyone hosting third-party generation must implement and maintain safeguards against infringing outputs.

Together these make concrete the difference between open weights and open source. The weights can be downloaded and run, but use is not unconditional: past a revenue threshold the licence is renegotiated, and liability is partly transferred to the user.

Why it matters

Music generation is more fragile than text or image on the copyright side. A melody resembling an existing work can carry legal consequences, and that risk sits with whoever runs the model. The licence's safeguard requirement points exactly there — MiniMax is not keeping the risk.

The real significance of open weights here is control. On a closed music service, what is produced and how it is filtered stay invisible; running it on your own server means you set both the output and the filter. A full five-minute song emerging in one generation is the technical threshold that makes that control usable in practice.