Deploy MiniMax Music 3

MiniMax Music 3 generates complete songs with vocals and accompaniment from lyrics and a music description, with support for songs up to approximately five minutes. It outputs 32 kHz, 16-bit stereo WAV audio.

Open release and license

Resource	Description
Model repository	`MiniMaxAI/MiniMax-Music3`
Inference code and tools	MiniMax-AI/MiniMax-Music3
License	MiniMax-Music3 Community License
SGLang-Omni Cookbook	MiniMax Music 3

The license includes product attribution, commercial revenue thresholds, safeguard, intellectual-property, and Acceptable Use Policy requirements. Read the complete license and implement all applicable terms before offering a generation service.

Input and output

Field	Description
`input`	Required lyrics; supports structure tags such as `[Verse]`, `[Chorus]`, `[Bridge]`, and `[Outro]`
`instructions`	Required description of genre, instruments, tempo, mood, vocals, and production
`seed`	Non-negative integer; fixed lyrics, description, length, and seed reproduce the same output
`max_new_tokens`	Audio-frame cap at 25 frames per second; the model may emit an end token earlier
Output	Non-streaming 32 kHz stereo WAV

Put each structure tag on its own line. For example, use [Verse]\nWalking down the street. Lyrics placed on the same line as a tag can be removed during normalization.

Install SGLang-Omni

The current SGLang-Omni Cookbook installs from source:

git clone https://github.com/sgl-project/sglang-omni.git
cd sglang-omni

uv venv .venv -p 3.12
source .venv/bin/activate
uv pip install -v -e .

Start the server

Single GPU

Colocate the autoregressive and acoustic stages on one GPU:

CUDA_VISIBLE_DEVICES=0 sgl-omni serve \
  --model-path MiniMaxAI/MiniMax-Music3 \
  --port 8000

Two GPUs

Place the autoregressive stage on the first GPU and the DiT/audio-decoder stage on the second GPU:

CUDA_VISIBLE_DEVICES=0,1 sgl-omni serve \
  --model-path MiniMaxAI/MiniMax-Music3 \
  --port 8000

The reference outputs in the SGLang-Omni Cookbook were generated on one H200 with the documented defaults. The single- and two-GPU commands describe stage placement and do not imply validation on every GPU memory capacity.

Generate your first song

curl -X POST http://localhost:8000/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MiniMaxAI/MiniMax-Music3",
    "input": "[Verse]\nMorning light across the floor\nThe quiet streets begin to wake\n[Chorus]\nStep into another day\nLet the music lead the way",
    "instructions": "Warm acoustic pop at 96 BPM with female vocals, fingerpicked guitar, soft piano, and drums entering gradually in a wide final chorus.",
    "seed": 7,
    "max_new_tokens": 750,
    "response_format": "wav",
    "stream": false
  }' \
  --output minimax_music3.wav

750 frames cap the output at approximately 30 seconds. max_new_tokens is a maximum, not a target. Generation ends earlier when the model emits the end-of-audio token.

Use the OpenAI SDK

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")

with client.audio.speech.with_streaming_response.create(
    model="MiniMaxAI/MiniMax-Music3",
    voice="default",
    input="[Verse]\nMorning light across the floor\n[Chorus]\nWe begin again",
    instructions="Warm acoustic pop at 96 BPM with female vocals, fingerpicked guitar, and soft piano.",
    response_format="wav",
    extra_body={"seed": 7, "max_new_tokens": 750, "stream": False},
) as response:
    response.stream_to_file("minimax_music3.wav")

The SDK streaming-response helper saves the HTTP response incrementally. Model generation itself remains non-streaming, and stream must be false.

Request limits

Both input and instructions must be non-empty.
The tokenized text prompt is limited to 5,000 tokens.
max_new_tokens is limited to 9,000 audio frames; the model may finish naturally before the cap.
Only response_format="wav" and stream=false are currently supported.
Do not send temperature, top_p, top_k, repetition_penalty, or reference-voice parameters; SGLang-Omni rejects them.
voice does not select a singer. Describe the vocals in instructions.

Capacity planning

Music 3 uses two generation stages and classifier-free guidance in both. Each request occupies two KV-cache rows in the autoregressive engine. When increasing concurrency, budget the KV cache for twice the number of sequences instead of using only the request count. To raise the concurrent-request limit:

CUDA_VISIBLE_DEVICES=0,1 sgl-omni serve \
  --model-path MiniMaxAI/MiniMax-Music3 \
  --port 8000 \
  --minimax_music3_ar.engine.max_running_requests 32

Validate lyrics and the music description with a short clip before rendering the complete song to reduce iteration cost.

Resources

SGLang-Omni MiniMax Music 3 Cookbook

Open the complete request contract, concurrency notes, and reproducible examples.

MiniMax Music 3 model repository

Open the model card, weights, architecture, and license.

Get started

Use Guides

FAQs

Terms & Policy

Open release and license

Input and output

Install SGLang-Omni

Start the server

Single GPU

Two GPUs

Generate your first song

Use the OpenAI SDK

Request limits

Capacity planning

Resources

SGLang-Omni MiniMax Music 3 Cookbook

MiniMax Music 3 model repository

​Open release and license

​Input and output

​Install SGLang-Omni

​Start the server

​Single GPU

​Two GPUs

​Generate your first song

​Use the OpenAI SDK

​Request limits

​Capacity planning

​Resources

SGLang-Omni MiniMax Music 3 Cookbook

MiniMax Music 3 model repository

Open release and license

Input and output

Install SGLang-Omni

Start the server

Single GPU

Two GPUs

Generate your first song

Use the OpenAI SDK

Request limits

Capacity planning

Resources