MiniMax · Music

MiniMax Music 3

Commercial use: conditions applyMiniMax-Music3 Community License

The heaviest listed model. The other listed entries are one GGUF file each; this one is a directory from which audio.cpp loads five components (condition encoder, language model, flow transformer, RVQ depth decoder, vocoder) plus tokenizer and config JSON. Commercial use is permitted, but it is capped by revenue and you must display the model's name in your product's interface.

MiniMaxAI/MiniMax-Music3 on Hugging Face audio.cpp

The licence

Summarized from the linked source documents, last checked 25 August 2026. Every claim links to its source; anything unconfirmed is marked unconfirmed.

MiniMax-Music3 Community License

Commercial use: conditions apply

Commercial use is permitted with three conditions: a revenue cap, a mandatory attribution notice in your product's UI, and safeguards if you offer it as a hosted service. Read it before you ship; the attribution duty is easy to miss.

Conditions

Prior written authorisation from MiniMax is required "if the aggregate yearly revenue generated from such products and services … exceeds 20 million US dollars".
Attribution: "You shall prominently display ‘MiniMax-Music3’ on the user interface of commercial product or service that uses the Software."
If you host it for others, you must implement "reasonable and proportionate technical and organizational safeguards" against violating uses and outputs.

What about the audio you generate?

The licence makes no explicit grant or assignment of output ownership. Its Acceptable Use Policy does restrict output: no IP-infringing content, no disinformation, and content published publicly must be clearly labelled as machine-generated. The Music 3 licence contains no explicit “MiniMax claims no rights over outputs” sentence. That language is in MiniMax's H3 licence; do not assume it applies here.

Correction to the shipped catalog: The shipped catalog entry says the licence was "not read" and asks you to treat it as conditional. It has now been read, and it is conditional: the conditions are the $20M revenue cap, the UI attribution notice, and hosted-service safeguards.

Based on the licence text as of the date below. Licences change; the links carry the authoritative version.

Specs

Transcribed from the shipped catalog, audiocpp_models_catalog.json — that file is authoritative if a number here disagrees with it.

Vendor: MiniMax
Catalog id: audiocpp-minimax-music3
audio.cpp loader family: minimax_music3
Can do: Generation, Lyrics
Weights repo: audio-cpp/MiniMax-Music3-GGUF
Weights (a package): config.json, config/**, tokenizer/**, language_model_q4_0.gguf, transformer_q4_0.gguf, rvq_depth_decoder_q8_0.gguf, condition_encoder.gguf, vocoder.gguf
Account needed to download: No
Max clip length: 120s
Sample rate: 44.1kHz stereo
Approx. VRAM: 10.0 GB
Download: ~8.5 GB
In the picker by default: Yes
Run end-to-end here: Yes

Getting it

There is no per-model install. One native binary runs all of these, so getting a model is a download of its weights and nothing else — no virtualenv, no torch pin, no CUDA wheel to match against your driver.

From the CLI

soundbox model install audiocpp
soundbox pull audiocpp-minimax-music3

The binary is fetched from audio.cpp's GitHub releases the first time you ask for it and cached under your own data directory. Soundbox ships no native binaries and vendors none.
Weights land in <models dir>/audiocpp/audiocpp-minimax-music3/ rather than in the shared Hugging Face cache, so a download stays scoped to the model you asked for.

Worth knowing

It will not run without lyrics. When you give none, Soundbox supplies an explicit [instrumental] marker so the run does not fail.
Component filenames are compiled into the audio.cpp binary and checked before any option is parsed, so no --session-option or --load-option can redirect them. The catalog entry mirrors audio.cpp 0.6.1's minimax_music3_q4_0 package exactly, mixed quantisation and all.
~10 GB of VRAM. On a modest card this is the listed entry most likely to fall back to CPU, which makes generation far slower.

Sources

Primary sources: the model's own repository, the project it came from, and the conversion audio.cpp actually downloads. Never invented — an entry with nothing findable says so instead.

Licence details on these pages are summaries of the linked source documents, last checked 25 August 2026. They are not legal advice, and Soundbox is not affiliated with any model's authors — weights download from Hugging Face under the user's own account and terms. Company and model names identify whose work it is, and belong to their owners. The linked licence text is authoritative; commercial use should be checked against it, per model, before anything ships.