vanch007/MiniMax-Music3-MLX-8bit

MiniMax Music 3 MLX 8-bit

Native Apple MLX checkpoint for MiniMaxAI/MiniMax-Music3, converted with selective affine 8-bit quantization and BF16 exceptions.

Runtime source and installation instructions: vanch007/mlx-minimax-music3.

Use

python -m pip install "mlx-minimax-music3[server] @ git+https://github.com/vanch007/mlx-minimax-music3.git@v0.1.0"

mlx-minimax-music3 generate \
  --model vanch007/MiniMax-Music3-MLX-8bit \
  --prompt "Warm acoustic pop with intimate female vocals and fingerpicked guitar." \
  --lyrics $'[verse]\nMorning light across the sea\n[chorus]\nStay here and sing this song with me' \
  --duration 10 \
  --seed 7 \
  --steps 30 \
  --output song.wav

Apple silicon and Metal are required. The production runtime does not use PyTorch.

Checkpoint

• Weight size: 14,167,660,156 bytes (13.195 GiB)
• Components: Qwen3 language model, RVQ depth decoder, condition encoder, flow transformer, stereo vocoder
• Output: 44.1 kHz stereo waveform
• Weight shards: 12
• MLX tensors: 1,978
• Quantization: affine 8-bit, group size 64

Quantized modules:

• Language-model attention and MLP matrices
• RVQ depth-decoder attention and MLP matrices
• Flow-transformer projections, attention, and feed-forward matrices

BF16 modules and parameters:

• Token, codebook, and positional embeddings
• Language-model and RVQ output heads
• Normalizations and biases
• Condition encoder
• Timestep and Fourier projections
• Convolutions and complete vocoder

Each component directory includes a conversion manifest with per-shard SHA-256 values, source-shard coverage, tensor names, and remote verification metadata.

Provenance

Item	Revision
Source model	MiniMaxAI/MiniMax-Music3@c2509fd6b60d1ae169cd1df27f78a53174ba17e8
Diffusers reference	huggingface/diffusers@c6da9936e4bda83107943a16eb8682e9a37d8527
MLX runtime	vanch007/mlx-minimax-music3@v0.1.0

The full provenance record is in source_manifest.json.

Validation

The complete repository was downloaded from its fixed Hugging Face commit and passed strict shard, SHA-256, tensor-name, and finite-value checks.

Fixed-input component comparisons against the official BF16 PyTorch reference produced:

Component	Cosine similarity	Relative RMSE
Language model	0.999822	0.025001
RVQ depth decoder	0.999916	0.012956
Condition encoder	0.999990	0.004654
Flow transformer	0.998990	0.045071
Vocoder	0.999877	0.015715

A local Apple M3 Max release run generated 10 seconds of audio using 250 autoregressive frames, two overlapping chunks, and 30 flow steps. The result was a finite, non-silent, unclipped 44.1 kHz stereo PCM WAV lasting 9.996 seconds. Peak MLX memory was 23.34 GiB.

These checks establish implementation and numerical alignment within the recorded quantized tolerances. Human listening review is pending; no claim of perceptual quality parity with the source release is made.

Limitations

• Apple silicon only
• One song per generation call
• No streaming output
• No training or fine-tuning support
• Generated audio may end before the requested maximum duration
• Lower-memory Macs have not been validated for this checkpoint

License

This converted checkpoint remains subject to the MiniMax-Music3 Community License, including its attribution, acceptable-use, safeguard, and commercial terms.