cowWhySo/MiniMax-Music3-TurboSpark-4bit-8step

MiniMax Music 3 TurboSpark, 4-bit, 8-step

A TurboSpark-compatible MiniMax Music 3 checkpoint for local Apple Silicon music generation. It combines the pinned mlx-community/MiniMax-Music3-4bit checkpoint with the community 8-step Turbo LoRA. The adapter is merged into the model's target matrices as FP16; the untouched affine INT4 group-64 linears remain quantized.

TurboSpark's native Rust/Metal runtime reads the model variant metadata in config.json and selects 8-step Euler sampling when the caller leaves the step count unspecified. The TurboSpark macOS app installs this exact repository revision through its pinned audio catalog. This is a merged checkpoint for the app, not an adapter-only download.

Provenance and prior work

The 4-bit base is mlx-community/MiniMax-Music3-4bit, pinned at commit c7ea32923b245fe5afc22d740a1936ad2ac590f3.

The adapter comes from guillaume127/MiniMax-Music-3-Turbo-FP8, specifically minimax_music3_turbo_lora_8step.safetensors. Its prior-art implementation and training details are in Guillaume-127/Minimax-music-3-Turbo-8-steps. TurboSpark did not train this adapter. Adapter SHA-256: 9722bb8da08da3912a2b6b8fdb173b9d7d4ce01a68818114a8218cfe25d8545a. The tested adapter-only MLX package is cowWhySo/MiniMax-Music3-Turbo-MLX-LoRA.

The rank-64, alpha-64 adapter targets the DiT in all 36 transformer blocks. Its fused QKV up-projection is split across Q, K, and V, yielding 216 residual projections. TurboSpark's merge probe covered all 216 targets; the report is in merge-report.json.

Use with TurboSpark

Install the model through TurboSpark's MiniMax Music 3 model picker. The app uses its immutable catalog pin and runs the native Rust/Metal backend. No Python or MLX runtime is needed in TurboSpark.

The project is whit3rabbit/turbospark. The checkpoint's metadata selects 8-step Euler with reference guidance. An explicit step count supplied by the caller still overrides that recommendation.

Evaluation status

This is an experimental checkpoint. One blind 15-second vocal comparison at seed 23 was judged similar to the 30-step reference; the 8-step no-adapter control was judged noticeably thinner. The adapter run took 55.18 seconds of process wall time on the test machine. This single seed does not establish broad audio quality, long-form behavior, or a stable performance result. Evaluate lyrics, genres, seeds, and 30–120 second generations before treating it as production-qualified.

Model size

The two merged weight shards are 12,671,777,748 bytes in the model index. The full local repository is about 12.68 GB decimal. The 216 adapted target matrices are FP16; all unaffected affine INT4 group-64 linears remain packed.

License

The checkpoint is a derivative of MiniMax Music 3 and remains subject to the MiniMax Music 3 Community License, including its acceptable-use and commercial terms. The source adapter repository declares Apache-2.0 for its adapter. Review both sources and their conditions before use or redistribution. The original MiniMax license is included as LICENSE.