cowWhySo/MiniMax-Music3-Turbo-MLX-LoRA

MiniMax Music 3 Turbo LoRA for MLX, 4-bit base, 8 steps

This repository publishes the 189 MB rank-64 Turbo LoRA separately from the 4-bit base checkpoint. Downloading the adapter alone is small; generation still requires the pinned 4-bit base model and generate.py, which attaches the adapter during inference.

Provenance

The adapter is from guillaume127/MiniMax-Music-3-Turbo-FP8. The upstream prior-art code and distillation description are at Guillaume-127/Minimax-music-3-Turbo-8-steps. TurboSpark did not train the adapter. Its SHA-256 is 9722bb8da08da3912a2b6b8fdb173b9d7d4ce01a68818114a8218cfe25d8545a.

It has rank 64 and alpha 64, targeting four named projections in each of 36 DiT blocks. The fused QKV target has one shared down projection and separate Q, K, V up projections. generate.py splits that target and attaches six low-rank residual projections per block, for 216 projection branches total. The quantized base weights remain unchanged in this MLX path.

The tested base is mlx-community/MiniMax-Music3-4bit, revision c7ea32923b245fe5afc22d740a1936ad2ac590f3. The loader was exercised with MLX 0.32.4 and mlx-audio 0.4.8 from development commit 784b29e2691a93ca7483147d86f61859dfaa6296.

Download and run

python -m pip install mlx==0.32.4 numpy safetensors
python -m pip install "mlx-audio @ git+https://github.com/Blaizzy/mlx-audio.git@784b29e2691a93ca7483147d86f61859dfaa6296"

hf download mlx-community/MiniMax-Music3-4bit \
  --revision c7ea32923b245fe5afc22d740a1936ad2ac590f3 \
  --local-dir ./base
hf download cowWhySo/MiniMax-Music3-Turbo-MLX-LoRA \
  --revision 9eec39a91fef90ff86ea98a6514534a12df1a6f2 \
  --include minimax_music3_turbo_lora_8step.safetensors \
  --local-dir ./adapter

python generate.py \
  --model ./base \
  --adapter ./adapter/minimax_music3_turbo_lora_8step.safetensors \
  --caption "Warm intimate vocal, piano and soft strings" \
  --lyrics "[verse]\nMorning light across the room" \
  --duration 15 --steps 8 --seed 23 --output vocal.wav

generate.py is a research harness from the TurboSpark experiment, included so the MLX adapter path can be reproduced. It defaults to 8 steps. Lyrics and audio quality remain model-dependent.

Evaluation limits

The adapter loaded and completed one 15-second vocal generation in the experiment. One blind seed comparison favored Turbo over the 8-step no-adapter control and judged it similar to the 30-step reference. This is not a multi-seed or long-form quality qualification. See the TurboSpark merged checkpoint model card for the experiment's results and caveats.

License

The source Turbo adapter repository declares Apache-2.0. Because the adapter is used with and was derived for MiniMax Music 3, use of the base-plus-adapter system remains subject to the MiniMax Music 3 Community License, including its acceptable-use and commercial terms. The adapter-only file is not a standalone generator.