What MiniMax-Music3 can do
MiniMax-Music3 basics
MiniMax-Music3 is a music generation model whose weights MiniMax has published.
A whole song, up to five minutes
Give it lyrics and a detailed music description, and a complete song up to five minutes long comes out. Not a short loop, but a single track with everything from intro to outro. Over long durations, it is stated to maintain themes, vocal identity, and the progression of the arrangement.
"MiniMax Music 3 is a high-performance music generation model for creating complete songs up to five minutes long. Conditioned on lyrics and a detailed music description, it generates structurally coherent songs with expressive vocals, evolving arrangements, and stable long-form audio quality."(opening)／"MiniMax Music 3 combines an 8B Global LLM for long-range musical structure, a 0.6B Local LLM for frame-level acoustic detail, and a continuous hidden-state synthesis system based on Flow Matching and Flow-VAE. The model produces 32 kHz, 16-bit stereo WAV audio."(opening)／"The model maintains musical themes, rhythm, vocal identity, and arrangement progression across long sequences, enabling complete structures such as intro, verse, pre-chorus, chorus, bridge, instrumental break, and outro."(Complete Songs with Long-Range Coherence section) — from the MiniMax-Music3 model card
Directing it with structure tags and a music description
In the lyrics you can put structure tags such as [Intro], [Verse], [Chorus], [Bridge], and [Outro] at the start of a line to state explicitly where the chorus is. In the music description, you specify genre, BPM, key, emotional progression, vocal character, instrumentation, and spatial treatment in words. The project recommends a structured caption split into three sections.
"Lyrics define the words to be sung and may include explicit section tags such as `[Intro]`, `[Verse]`, `[Pre-Chorus]`, `[Chorus]`, `[Post-Chorus]`, `[Bridge]`, `[Instrumental]`, `[Solo]`, and `[Outro]`."／"Music description defines the musical style, emotional progression, vocal performance, instrumentation, arrangement, and production profile."／"For precise control, we recommend using a Structured Caption with three sections:"(all from the Fine-Grained Music Control section) — from the MiniMax-Music3 model card
How to use MiniMax-Music3
VRAM required
Layer streaming drops the requirement to 8GB, so a 24GB card is not required. Full precision and CPU offload differ by only about 2GB, making this effectively a choice between ~24GB and 8GB.
Inference is supported by SGLang-Omni, diffusers, and ComfyUI. The steps below use SGLang-Omni.
Running it with SGLang-Omni
First download the weights, then start it as a server. Generation uses the shared speech API endpoint: lyrics go in input, and the music description goes in instructions.
hf download MiniMaxAI/MiniMax-Music3 --local-dir /path/to/minimax_ttm
sgl-omni serve --model-path MiniMaxAI/MiniMax-Music3 --port 8000
max_new_tokens is an upper bound counted at 25 frames per second. Generation stops earlier if the model emits an end-of-audio token.
"MiniMax Music 3 is supported by [SGLang-Omni]"／"The service uses the shared speech API. Put the lyrics in `input` and the music description in `instructions`."／"`max_new_tokens` sets the maximum number of audio frames at 25 frames per second. Generation may finish before this limit when the model emits an end-of-audio token."／"The response is a 32 kHz, 16-bit stereo WAV file."(all from the How to Use section) — from the MiniMax-Music3 model card
Running it on an 8GB GPU
Full precision fits under 24GB, automatic CPU offloading brings it to about 22GB, and streaming the language model layer by layer fits it onto an 8GB video card. Generation is slower in the offloading configurations. For the practicalities of running LLMs locally, see making local LLMs practical for coding and the measured local GLM article.
"The full precision fits under 24GB of VRAM. With automatic CPU offloading, generation takes in ~22 GB; additionally streaming the language model layer by layer makes it fit even 8 GB video cards"(Low VRAM section) — from the MiniMax-Music3 model card
The license to check before you use it
Conditions for commercial use
The weights are public, but the license is not MIT.
Commercial terms in the community license
The license is the custom “MiniMax-Music3 COMMUNITY LICENSE.” If you build it into a commercial product or service, you are obliged to display “MiniMax-Music3” prominently in the user interface. On top of that, if aggregate yearly revenue across you and your affiliates exceeds 20 million US dollars, you must contact MiniMax and obtain separate prior written authorization. If you provide outputs to third parties, you are also required to implement and periodically review safeguards against infringing uses.
"You shall prominently display “MiniMax-Music3” on the user interface of commercial product or service that uses the Software."／"You shall obtain a separate, prior written authorization from MiniMax … if the aggregate yearly revenue generated from such products and services provided by you and/or any of your affiliates exceeds 20 million US dollars (or equivalent in other currencies)."(both from clause 3, Commercial Terms)／"you must, before making that product or service available and throughout its operation, implement, maintain, test, and periodically review reasonable and proportionate technical and organizational safeguards designed to prevent and mitigate access, uses, and outputs that violate this License"(clause 4) — from the MiniMax-Music3 COMMUNITY LICENSE
A realistic way to try it
Check the sound in a personal prototype, then re-read the license when you move toward commercialization. That order is the safe one. Holding on to the distinction between "the weights are public" and "you may use it commercially however you like" avoids having to rebuild later.
When you want to convert an English model card into markdown with its structure intact, the following tool can help.
