MiniMax Music 3

MiniMax Music 3 is an open music model from MiniMax that creates a full song with vocals from lyrics and a style description. Per the model card, a track can run up to 5 minutes; lyrics can carry structure labels such as [Verse] and [Chorus]. The output is stereo WAV, 32 kHz, 16-bit. Inside are language models of 8B and 0.6B parameters and a 2.4B flow-matching model. The card lists limits: CUDA only, no streaming mode, and a prompt limit of 5000 tokens. The weights are public on Hugging Face without any acceptance step and released on 2026-08-07. The license is the MiniMax-Music3 Community License (the license file in the model repository). Per the license text, a commercial product or service must prominently display "MiniMax-Music3" in its interface, and if yearly revenue from such products exceeds 20 million US dollars, prior written permission from MiniMax is needed; services for third parties must have safeguards against outputs that violate the license, and an Acceptable Use Policy is attached. The text has no territorial restrictions and no direct clause on rights to outputs. The authors offer an official demo Space on Hugging Face and a demo page with examples. The same developer's video model is in our catalog: MiniMax H3. Date of check: 2026-10-18.