MiniMax Music 3 — MLX

Precision-preserving MLX-native layout conversion of MiniMaxAI/MiniMax-Music3 for local inference on Apple silicon. It is designed for use with mlx-minimax-music3, the independent pure-MLX inference project and Python package that generates complete stereo music from lyrics and a structured music caption without PyTorch, CUDA, or a cloud API at inference time.

This is a format and tensor-layout conversion. It is not trained, fine-tuned, merged, or quantized, and it does not claim authorship of the underlying model. MiniMax developed and released MiniMax Music 3; App Automaton converted the published checkpoint for the independent MLX runtime.

Checkpoint contents

Component	Stored dtype	Size	Role
Global language model	BF16	15.99 GiB	Long-range structure and semantic music tokens
RVQ depth decoder	BF16	1.20 GiB	Seven residual acoustic codebooks
Condition encoder	FP32	0.09 GiB	Continuous hidden-state fusion
Flow transformer	FP32	9.06 GiB	Flow-matching acoustic synthesis
Vocoder	FP32	0.20 GiB	Stereo waveform decode
Tokenizer, scheduler, and metadata	—	0.01 GiB	Prompting and checkpoint contract

The complete checkpoint is 26.56 GiB (28.52 GB decimal). The repository contains only the dense profile. It does not contain selective-q8 or persistent FP16 derivatives.

Conversion contract

The conversion is pinned to official source revision fbdf52fbaaca799592917417eb05f1899f1255ec. Its manifest.json records the source revision, mapping version, component file sizes, tensor counts, dtypes, and SHA-256 digests.

Qwen3 and RVQ tensors retain their published BF16 values.
Condition, flow, and vocoder tensors retain their published FP32 values.
Convolution kernels are transposed into the channels-last layout expected by MLX.
Vocoder weight normalization is folded into the stored convolution weights.
No tensor is downcast or quantized.

The converter reads and writes SafeTensors directly through MLX. PyTorch is not part of conversion or runtime inference.

Use with MLX

Install the current package from mlx-minimax-music3 on PyPI:

uv add --prerelease=allow mlx-minimax-music3

Download this checkpoint into a local weight directory:

hf download appautomaton/MiniMax-Music3-MLX \
  --local-dir weights/mlx-dense/MiniMax-Music3

Generate one minute of instrumental melodic techno:

from mlx_minimax_music3 import (
    GenerationRequest,
    Music3Pipeline,
    instrumental_lyrics,
)

pipeline = Music3Pipeline("weights/mlx-dense/MiniMax-Music3")
result = pipeline.generate(
    GenerationRequest(
        caption=(
            "Global Metadata: melodic techno, 128 BPM, A minor, nocturnal and "
            "cinematic, gradually rising energy. Vocal Details: instrumental, "
            "no vocals. Arrangement: deep rounded kick, warm sub-bass, crisp "
            "hats, syncopated percussion, analog arpeggiator, evolving pads, "
            "a glassy bell motif, controlled builds, and a spacious final drop."
        ),
        lyrics=instrumental_lyrics(
            "intro", "groove", "build", "drop", "breakdown", "outro"
        ),
        audio_duration=60.0,
        seed=7,
    ),
    output="outputs/melodic-techno.wav",
)

print(result.metadata.stage_timings)
print(result.metadata.memory_reports)

audio_duration is a ceiling because the model may emit its end token earlier. Set min_audio_duration when a minimum frame count is required. The default checkpoint path keeps the official mixed precision: BF16 autoregressive models and FP32 acoustic models.

Runtime behavior

The runtime loads one stage at a time. Autoregressive models are released before the flow transformer is loaded, and acoustic models are released before final waveform decoding. This bounds unified-memory residency and avoids retaining the entire checkpoint in memory at once.

The current runtime writes native 44.1 kHz stereo PCM16 WAV. The official serving profile resamples its output to 32 kHz; reference-output parity for that final profile remains in progress.

Validation status

This is an alpha release. The dense checkpoint has passed:

strict tensor-name, shape, dtype, and shard-index validation;
tensor-by-tensor conversion checks against the pinned source;
complete checkpoint manifest digest verification;
weightless golden regression tests for dense loading and inference; and
end-to-end local generation, including a three-minute default-FP32 run.

On an Apple M5 Max with 128 GB unified memory, the three-minute validation run took 18 minutes 16 seconds, peaked at approximately 19.93 GiB of process memory, and did not increase swap usage. This is one machine-specific observation, not a portable performance guarantee.

Listening validation across more prompts and seeds, long-form quality parity, and the reference 32 kHz output profile are still in progress.

Intended use and limitations

This checkpoint is intended for local research, development, and music generation with the MLX runtime on Apple silicon.

Prompt controls such as tempo, key, instrumentation, lyrics, and structure are generative guidance rather than strict symbolic guarantees.
Outputs can contain artifacts, incorrect words, unexpected structure, or content that does not follow every requested attribute.
Users are responsible for evaluating generated content, respecting applicable rights, and complying with the model license and acceptable-use policy.
The checkpoint is not an official MiniMax MLX release, and this project is not affiliated with or endorsed by MiniMax.

For the original architecture description, prompt guidance, examples, and model limitations, read the MiniMaxAI/MiniMax-Music3 model card.

License

The converted checkpoint remains governed by the included MiniMax-Music3 Community License, including its attribution, acceptable-use, safeguards, and commercial terms. Review that license before downloading, redistributing, or deploying the model.

The mlx-minimax-music3 runtime code is separately licensed under MIT.

appautomaton
/

MiniMax-Music3-MLX