ACE-Step 1.5 sft gets an 8-bit Mac build under MIT

A converted build of the ACE-Step 1.5 music model went up for Apple Silicon on September 26, 2026, and its author's notes say the obvious way to run it produces noise instead of singing.

The build is roman220220/ACE-Step1.5-sft-MLX-8bit, an 8-bit conversion of the ACE-Step team's sft checkpoint into the MLX format used on Apple Silicon Macs. It makes songs with sung, intelligible lyrics from a style description at 48 kHz stereo. The Hub records the release as MIT, the same licence as the base model, and the author states the base was trained on licensed and royalty-free material. The upstream card adds that the music it generates can be used for commercial purposes.

What it costs to run is the other half. The diffusion transformer is 2.7 GB, the full download including the Qwen3-Embedding text encoder and the VAE is about 5.6 GB, and the card reports roughly 49 seconds of compute for 30 seconds of music on an M5 at 50 steps, peaking around 8.6 GB of memory. The card also lists a full-precision (bf16) build at about 36 seconds and 10.6 GB, and a 4-bit GPTQ build at the same 36 seconds and 7.5 GB, which the author recommends for a 16 GB Mac. Quantizing to 8-bit makes it slower here, not faster, and the author puts that down to the 8-bit matrix multiply.

Lyric clarity is measured rather than asserted. The author scored intelligibility by transcribing the output with Whisper and comparing it against the lyrics that were asked for, across three songs and four seeds each. The faster turbo model, as a 4-bit build with a 1.7B planner, scores a mean word error rate of 0.56. The full-precision sft build scores 0.22, this 8-bit conversion 0.24 and the 4-bit GPTQ build 0.29. With only twelve tracks, the card warns, differences of a few points are noise.

Run through the community mlx-audio toolkit's defaults, the sft model produces cacophony. The author found two causes. The first is that mlx-audio feeds audio codes from its language-model planner into the music model, and that breaks the sft build, which does not need a planner. Turning the planner off with use_lm=False makes it sing and saves about fifteen seconds a track. The second is the guidance branch, where mlx-audio encodes all-zero text: the card's patch swaps in the model's own trained blank input, which is what the official pipeline uses. In the findings log, that took mean word error rate from 0.28 to 0.22 on the full-precision build.

For a creator, the practical upshot is a music generator whose licence permits selling the result, running on hardware already on the desk, with the exact commands and both fixes published alongside the weights.

The weights are on Hugging Face, and the findings log with the measurements is linked from the card.