mlx-community/MiniMax-Music3-mxfp8

MiniMax Music 3 · MLX MXFP8

Native MLX MXFP8 weights for MiniMaxAI/MiniMax-Music3, converted for lyric-conditioned song generation on Apple Silicon with mlx-audio. This is the recommended quantized variant when lyric fidelity matters.

Community conversion, not an official MiniMax release. All model credit goes to MiniMax. Review the original model card and license before use.

Other MLX variants: BF16 · 8-bit · 6-bit · 4-bit · MXFP4 (experimental) · NVFP4 (experimental)

Install

MiniMax Music 3 support is proposed in Blaizzy/mlx-audio#888. Until that PR is released, install the tested commit directly:

python -m pip install "mlx-audio @ git+https://github.com/pinglin/mlx-audio.git@c2fa486ddf98ea8268c227b6d624b2f9ba00dbab"

Generate

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

from mlx_audio.music import load

model = load("mlx-community/MiniMax-Music3-mxfp8")
result = next(
    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)  # stereo, 44100 Hz

Lyrics are required by the checkpoint contract. Use [instrumental] explicitly for instrumental generation. Duration is a requested upper bound: the autoregressive stage may emit its end token early. Style, tempo, instrument, and vocal controls are probabilistic rather than strict.

Quantization and verification

• MXFP8, group size 32: E4M3 linear weights with E8M0 group scales.
• Large linears in the global language model, RVQ depth decoder, and flow transformer are quantized. Embeddings, output heads, convolutions, condition encoder, and vocoder remain dense for fidelity.
• Approximately 13 GB on disk, versus 27 GB for the BF16 conversion.
• Strict loading and real generation produced finite 44.1 kHz stereo audio.
• Controlled lyric checks across two seeds retained substantially more requested text than MXFP4 and stayed close to the dense BF16 baseline. For example, the 24-second seed-113 chorus preserved nearly every requested phrase, while MXFP4 collapsed to a short unrelated fragment.
• The mlx-audio regression suite passed 1,742 tests with 34 expected skips; the focused music, converter, and registry suite passed 43 tests and 3 subtests.

Converted with mlx-audio 0.4.8 development commit c2fa486 and MLX 0.31.2.

License

The weights remain subject to the MiniMax-Music3 Community License, including its acceptable-use and commercial terms. The full license text is included in this repository.