MiniMax Music 3¶

MiniMax Music 3 is a multilingual song-generation model that combines a Qwen3 autoregressive planner, an RVQ depth decoder, a flow-matching transformer, and a stereo 44.1 kHz vocoder. The original checkpoint is MiniMaxAI/MiniMax-Music3.

Models¶

Model	Recommendation
`mlx-community/MiniMax-Music3-bf16`	Dense reference conversion
`mlx-community/MiniMax-Music3-8bit`	Affine 8-bit conversion
`mlx-community/MiniMax-Music3-6bit`	Affine 6-bit conversion
`mlx-community/MiniMax-Music3-4bit`	Affine 4-bit conversion
`mlx-community/MiniMax-Music3-mxfp8`	Recommended balance of memory and lyric fidelity
`mlx-community/MiniMax-Music3-mxfp4`	Experimental lower-memory conversion
`mlx-community/MiniMax-Music3-nvfp4`	Experimental NVFP4 conversion

Choose MXFP8 when lyric fidelity matters

MXFP4 uses less memory, but may alter or omit more requested words than BF16 or MXFP8.

Generate¶

CLIPython

python -m mlx_audio.music.generate \
    --model mlx-community/MiniMax-Music3-mxfp8 \
    --caption "Warm acoustic pop, 96 BPM, intimate female vocal" \
    --lyrics $'[verse]\nMorning light across the room\n[chorus]\nSing with me' \
    --duration 30 \
    --steps 30 \
    --seed 7 \
    --output song.wav

For longer lyrics, pass a UTF-8 file with --lyrics-file lyrics.txt.

from mlx_audio.music import load

model = load("mlx-community/MiniMax-Music3-mxfp8")

for result in model.generate(
    text="Warm acoustic pop, 96 BPM, intimate female vocal",
    lyrics="[verse]\nMorning light across the room\n[chorus]\nSing with me",
    duration=30,
    steps=30,
    seed=7,
):
    print(result.audio.shape, result.sample_rate)

Place section tags such as [verse], [chorus], and [bridge] on their own lines. Lyrics are required by the checkpoint contract; use [instrumental] explicitly for instrumental generation.

Generation supports up to 360 requested seconds, but the autoregressive model may emit its end token earlier. Caption controls such as tempo and instrumentation are probabilistic rather than strict.

Manual conversion¶

The general converter reads the official modular checkpoint directly:

python -m mlx_audio.convert \
    --hf-path MiniMaxAI/MiniMax-Music3 \
    --mlx-path ./MiniMax-Music3-mxfp8 \
    --quantize \
    --q-mode mxfp8

The MLX implementation is adapted from mikolaj92/minimax-music3-mlx under Apache-2.0. The model weights and tokenizer use the MiniMax-Music3 Community License.