MiniMax · Music

MiniMax H3

Commercial use: conditions applyMiniMax H3 Community License Agreement

It needs roughly 36 GB of VRAM, an order of magnitude more than the Stable Audio entries. Upstream it is a very large omni-modal model that generates video with native audio; it is catalogued here for its audio side. Its licence is the only one in this catalog with a geographic exclusion, and the excluded regions are the US, EU, UK and South Korea.

MiniMaxAI/MiniMax-H3 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 H3 Community License Agreement

Commercial use: conditions apply

Commercial use is permitted under a revenue cap and an attribution duty, like MiniMax Music 3. The larger restriction is territorial: the licence's “Applicable Territory” is the world minus the EU, the UK, South Korea and the USA. In those regions you must apply to MiniMax separately before any use.

Conditions

Territory: the EU, UK, South Korea and USA are excluded from the standard grant; the model card carries a separate application form for those regions.
Prior written authorisation required above US$20M yearly revenue.
Attribution: you must "prominently display ‘MiniMax H3’ on the user interface".
MiniMax's original weights are gated behind the Community License Agreement on Hugging Face. Soundbox downloads an ungated GGUF conversion; the agreement still applies to it.

What about the audio you generate?

This licence covers output, in your favour: MiniMax “claims no rights over the Outputs you generate”, while making you “entirely responsible for the Outputs and any subsequent use thereof”. Two restrictions ride along: outputs may not be used to improve competing AI models, and machine-generated content must be disclosed in public settings.

Correction to the shipped catalog: The shipped catalog lists this licence as "Unknown". It is the MiniMax H3 Community License, the most restrictive licence in this catalog for anyone in the US, UK, EU or South Korea.

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-h3
audio.cpp loader family: minimax_h3
Can do: Generation
Weights repo: audio-cpp/audio.cpp-gguf
Weights (a package): MiniMax-H3-Q4-GGUF/configuration.json, MiniMax-H3-Q4-GGUF/FL2VA/processor/**, MiniMax-H3-Q4-GGUF/text_encoder_q4_k.gguf, MiniMax-H3-Q4-GGUF/dit.gguf, MiniMax-H3-Q4-GGUF/audio_vae_folded_f16.gguf, MiniMax-H3-Q4-GGUF/video_vae.gguf
Account needed to download: No
Max clip length: 60s
Sample rate: 44.1kHz stereo
Approx. VRAM: 35.9 GB
Download: ~8.45 GB of a 58 GB repo
In the picker by default: No — opt in from Settings
Run end-to-end here: No — catalogued, not tested here

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-h3

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-h3/ rather than in the shared Hugging Face cache, so a download stays scoped to the model you asked for.
This one is opt-in: switch it on in the Models page's “Enable more models” section before it appears in the picker.

Worth knowing

A package model living under a prefix in the shared repo, so the catalog entry needs weights_root to point --model at the right subdirectory.
The repo it lives in is 58 GB; the entry pulls about 8.45 GB of it. That bounded download is why weights_glob accepts a list.
~36 GB of VRAM. On anything smaller this falls back to CPU, where it is not practically usable.

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.