---
library_name: sglang-omni
pipeline_tag: text-to-audio
tags:
  - music-generation
  - text-to-music
  - pytorch
  - sglang-omni
---

# MiniMax Music 3

MiniMax Music 3 is a non-streaming music generation model conditioned on lyrics
and a natural-language music description. Inference is implemented in
SGLang-Omni with PyTorch and uses two CUDA GPUs:

- GPU 0: Qwen3 + eight-codebook RVQ autoregressive generation;
- GPU 1: flow-matching DIT + DAV waveform decoding.

The generated audio is returned as a 32 kHz, 16-bit stereo WAV file.

## Model files

```text
minimax_ttm/
  config.json
  flowmatching_vae.pth
  dav.pth
  qwen_7B/
    qwen_7B/
      config.json
      model.safetensors.index.json
      model-*.safetensors
    qwen3-8B-tokenizer-music/
      tokenizer.json
      tokenizer_config.json
      ...
  assets/
    minimax_ttm.wav
```

## Usage

Download the repository to a local directory before starting the service:

```bash
hf download MiniMaxAI/MiniMax-Music3 --local-dir /path/to/minimax_ttm
```

From the SGLang-Omni repository root:

```bash
python -m sglang_omni.cli serve \
  --config examples/configs/minimax_ttm.yaml \
  --model-path /path/to/minimax_ttm \
  --host 127.0.0.1 \
  --port 8000
```

Send a non-streaming request through the shared speech API. `input` contains
the lyrics and `instructions` contains the music description.

```bash
curl http://127.0.0.1:8000/v1/audio/speech \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "minimax_ttm",
    "input": "[Verse]\nHello from SGLang Omni",
    "instructions": "A bright piano pop song with a warm female vocal",
    "response_format": "wav",
    "seed": 7,
    "max_new_tokens": 9000,
    "stream": false
  }' \
  --output minimax_ttm.wav
```

An example generated with this checkpoint is available at
[`assets/minimax_ttm.wav`](assets/minimax_ttm.wav).

## Limitations

- Inference requires two CUDA GPUs.
- Only non-streaming generation is supported.
- The tokenized text prompt is limited to 5000 tokens.
- Audio generation is limited to 9000 acoustic frames.