News · 2026-08-13

MiniMax released a five-minute song model with a catch in the licence

MiniMax published the weights for MiniMax Music 3, a model that generates complete songs up to five minutes long with sung vocals, arrangement, and 32 kHz 16-bit stereo output. It is a genuine open-weight release on Hugging Face rather than a hosted demo. The licence, however, is not an open-source licence: commercial use is permitted by default but requires the name "MiniMax-Music3" displayed prominently in the product's interface, and any organisation above 20 million US dollars in aggregate yearly revenue needs prior written authorisation.

Key facts

• Generates full songs up to five minutes with intro, verse, pre-chorus, chorus, bridge, instrumental break, and outro structure, at 32 kHz 16-bit stereo.
• Architecture: an 8B global model initialised from Qwen3-8B for long-range structure, a 0.6B local model for frame-level acoustic detail, a 2.4B flow-matching synthesis stage, and a 123M Flow-VAE decoder.
• Licence: the MiniMax Community License allows commercial use with a mandatory on-screen attribution and a written-authorisation requirement above 20 million US dollars in yearly revenue.
• Primary source: the MiniMax Music 3 model card on Hugging Face.

The hard problem in music generation is not making thirty seconds sound good. It is making minute four remember minute one -- keeping the same vocalist, the same key, the same motif, while the arrangement actually develops rather than looping. Systems that generate audio in one pass tend to drift; the singer subtly changes, the theme dissolves.

MiniMax's answer is to split the job by timescale. A large 8-billion-parameter global model, initialised from Qwen3-8B, predicts only the first codebook frame by frame and carries the song's long-range semantic and structural progression -- the part that has to remember. A much smaller 0.6-billion-parameter local model fills in the remaining acoustic detail within each frame. The two are trained jointly. It is the division of labour between a composer holding the whole arc of a piece in mind and a session player producing the texture of each bar, and it is a sensible way to spend parameters: structure needs capacity and memory, timbre needs speed.

The audio itself is not decoded from discrete tokens in the usual way. The synthesis path fuses the final hidden states of both models, passes them through a 2.4-billion-parameter flow matching stage into a Flow-VAE latent, and decodes to waveform through a 123-million-parameter decoder -- deliberately skipping the discrete tokenizer at inference so that the continuous representation, which preserves more information about vocal articulation and instrumental texture, survives to the output. The tokenizer used in training stacks eight layers of residual vector quantization: one semantic codebook of 16,384 entries capturing musical structure, plus seven acoustic codebooks of 1,024 entries each for residual detail. The Flow-VAE architecture is adapted from MiniMax's speech model and retrained for the dynamic range of music.

Control is split into two fields, which is the part that matters for anyone actually using it. Lyrics carry the words plus explicit section tags -- [Intro], [Verse], [Pre-Chorus], [Chorus], [Bridge], [Instrumental], [Solo], [Outro] -- each on its own line. A separate music description carries style, emotional progression, vocal performance, instrumentation, and production profile, and MiniMax recommends a structured caption split into global metadata, vocal details, and arrangement. Generation runs through SGLang-Omni, Diffusers, or ComfyUI, using the same speech API shape with lyrics in the input field and the description in instructions.

The licence is where this release differs most from the surrounding conversation about open-weight models. Downloading and running the weights is unrestricted. Building a product on them is not: the attribution requirement puts MiniMax's name in your interface, and the revenue threshold means a successful product eventually has to come back and ask. That is a deliberate structure, not an oversight -- it captures the long tail of hobbyists and small builders for free while retaining a negotiating position with anyone who succeeds. "Open weights" and "open source" continue to be different things, and the difference is usually in a file most people never open.

Context makes the release sharper. Suno, the dominant commercial music generator, has been moving in the opposite direction, having already told users through its Warner Music Group partnership announcement that downloading songs would require a paid account with tier-specific monthly limits. A downloadable model that runs on your own hardware with no per-song meter is a direct answer to that, subject to the licence.

The honest caveats: MiniMax publishes no independent evaluation on the model card, so quality claims rest on the demos and on user reports rather than on comparative measurement, and the card includes its own limitations section. Hardware requirements for the full stack are non-trivial given a combined footprint above 11 billion parameters across four components. And the legal question hanging over every music generator -- what the training data was, and what that means for output you intend to sell -- is not addressed by publishing weights. Open weights answer where the model runs. They do not answer where it came from.

Primary source, verified: read the paper →

Ground Truth. (2026, August 13). MiniMax released a five-minute song model with a catch in the licence. Ground Truth. https://groundtruth.day/news/minimax-released-a-five-minute-song-model-with-a-catch-in-the-licence.html

@misc{groundtruth:minimax-released-a-five-minute-song-model-with-a-catch-in-the-licence,
  title  = {MiniMax released a five-minute song model with a catch in the licence},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {aug},
  url    = {https://groundtruth.day/news/minimax-released-a-five-minute-song-model-with-a-catch-in-the-licence.html}
}

Topics: open-weights · music · generative-audio · minimax · licensing