MiniMax Music 3 MLX
Collection
Apple Silicon MLX port of MiniMax Music 3: bf16 plus 8/6/4-bit quants. • 4 items • Updated
How to use mikoy92/MiniMax-Music3-MLX with MLX:
# Download the model from the Hub pip install huggingface_hub[hf_xet] huggingface-cli download --local-dir MiniMax-Music3-MLX mikoy92/MiniMax-Music3-MLX
Converted MiniMaxAI/MiniMax-Music3 weights for Apple Silicon via mikolaj92/minimax-music3-mlx.
This is a layout remap (PyTorch conv kernels → MLX, weight-norm fused, bf16). It is not a fine-tune.
All variants: MiniMax Music 3 MLX collection
| Precision | Repo | Size |
|---|---|---|
| bf16 | mikoy92/MiniMax-Music3-MLX | ~23.5 GB |
| 8-bit | mikoy92/MiniMax-Music3-MLX-8bit | ~12 GB |
| 6-bit | mikoy92/MiniMax-Music3-MLX-6bit | ~10 GB |
| 4-bit | mikoy92/MiniMax-Music3-MLX-4bit | ~7 GB |
git clone https://github.com/mikolaj92/minimax-music3-mlx
cd minimax-music3-mlx
uv venv --python 3.12 .venv
source .venv/bin/activate
uv pip install -e .
hf download mikoy92/MiniMax-Music3-MLX --local-dir ./weights/mlx
python -m minimax_music3_mlx.generate \
--weights ./weights/mlx \
--lyrics $'[verse]\n...\n[chorus]\n...' \
--prompt "Genre: acoustic pop. BPM: 96. Warm female vocal." \
--duration 30 \
--steps 30 \
--output song.wav
Weights remain under the MiniMax-Music3 Community License (Copyright 2026 MiniMax). Inference code in the GitHub repo is Apache-2.0.
Commercial products that ship the model must display “MiniMax-Music3”. Yearly revenue above USD 20M needs a separate MiniMax authorization (api@minimax.io).
Quantized
Base model
MiniMaxAI/MiniMax-Music3