MiniMax-Music3-MLX-Serve-8bit

MiniMax Music 3 for mlx-serve (8-bit)

MiniMax-Music3 converted for mlx-serve's native Zig + MLX engine. Full songs with sung lyrics at 44.1 kHz stereo, generated locally on Apple Silicon.

Quantization: affine 8-bit, group 64, on every real matmul (LLM, depth decoder, DiT, lm_head). Kept dense on purpose: the embedding tables (gather reads), the condition encoder and the whole vocoder (VAE-class precision). Worst per-tensor reconstruction error 1.48% RMS. 13 GB on disk instead of the 57 GB upstream repo (which ships its weights twice).

Engine parity against the fp32 reference on these exact weights: prefill cosine 0.9999, condition encoder 0.999999, DiT velocity 0.999, vocoder 1.000000. The autoregressive stage runs about 44 ms per frame on an M-series Mac, so a one-minute song takes roughly a minute of LLM time plus the diffusion pass.

Run it

Download MLX Core.app, open the Music tab and pick MiniMax Music 3. Style prompt + lyrics in, WAV out.

Over HTTP:

mlx-serve --serve
curl http://127.0.0.1:11234/v1/audio/music-generations \
  -H 'Content-Type: application/json' -o song.wav -d '{
    "model": "MiniMax-Music3-MLX-Serve-8bit",
    "prompt": "upbeat synthwave with driving bass and dreamy pads",
    "lyrics": "[verse]\nneon lights across the bay\n[chorus]\nwe run all night",
    "duration_seconds": 60
  }'

Lyrics are required (the model is lyric-conditioned) and structure tags like [verse] or [chorus] go on their own lines. duration_seconds (1-360) is an upper bound, the model may end the song earlier. ACE-Step style fields (bpm, keyscale, timesignature, vocal_language) do not exist on this model.

Rebuild from the upstream repo with scripts/convert_music3_weights.py.

Weights are covered by the MiniMax-Music3 Community License (see LICENSE, acceptable-use policy included as Exhibit A).