What is MiniMax Music? Full songs, section control and open-weight conditions
MiniMax Music creates music from descriptions and lyrics. Learn how Music 3.0 handles full-song structure, audio formats and local deployment, including community-license conditions and non-streaming limitations.
Meet this model
MiniMax Music is MiniMax’s family for creating music from descriptions and optional lyrics. Music 3.0 provides open weights for full-song structure and vocals, unlike MiniMax Speech’s spoken narration. Online apps, cloud APIs and local weights are separate usage routes and do not establish support across every version.
Describe how sections change across a complete song
Music 3.0 can generate complete songs up to about five minutes from lyrics and detailed descriptions. For a quiet verse leading into a bright chorus, specify arrangement and vocal changes in each section rather than only name a genre such as pop.
Section tags guide generation without strict guarantees for every beat, tempo or key. The official card acknowledges possible departures in lyrics and structure. For a short film, compare important cuts against the complete track before choosing trimming or regeneration.
Global structure and local acoustic detail are modeled separately
Music 3.0 combines global and local language modeling with continuous hidden-state audio synthesis. The global part addresses longer structure and the local part finer acoustic detail. This explains the design without establishing that every long song avoids repetition or pitch problems.
To check a chorus or vocal clarity, retain the entire track and input lyrics and identify the section where it deviates. Official samples do not establish pronunciation of a user’s Chinese proper names. This is documentation-based explanation without generated songs or long-form stability tests.
Local output is an audio file rather than a live performance stream
The official card lists 32 kHz, 16-bit stereo WAV output and specifies CUDA inference with non-streaming generation only. Do not describe those weights as real-time accompaniment on an ordinary phone or infer a streaming service from a playback example.
Use the actual sample rate when importing into an edit to avoid speed errors. For a loop, listen to the join between end and start. A sufficiently long song does not mean any segment loops seamlessly. No inference environment or device comparison was tested here.
The Music3 community license includes commercial conditions
Music3 weights use the MiniMax-Music3 Community License, including attribution and acceptable-use requirements. Commercial interfaces must prominently display MiniMax-Music3, and aggregate annual revenue from relevant products and services above USD 20 million requires separate prior written authorization. Consult the full license.
The license also includes safeguards for public generation services. Check online account terms separately; an open-weight license does not replace subscription or API rules. A model’s ability to create a song does not establish permission to use supplied lyrics or reference works.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.