ACE-Step 1.5 XL: MIT-licensed music generation that fits in 12GB

StepFun open-sourced ACE-Step 1.5 XL, a music-generation model built around a 4-billion-parameter DiT (diffusion transformer) decoder, the type of architecture used in several image and video generators, applied here to audio. It ships in three variants, and StepFun says the smallest fits on 12GB of VRAM using INT8 offloading, meaning part of the model runs at lower numeric precision to save memory, while full-quality generation needs 24GB.

The model carries an MIT license, and StepFun describes its training data as commercially safe, meaning teams can in principle generate and ship music from it inside a commercial product without the licensing uncertainty around music models trained on scraped or unlicensed audio. StepFun did not publish independent audio-quality benchmarks alongside the release, so how the output compares to closed music generators remains unverified here.