ACE-Step 1.5 XL brings local AI song generation to consumer GPUs

ACE-Step 1.5 XL is an open-source AI music generator you can run locally on your own machine. Feed it a prompt describing the musical vibe alongside your lyrics, and it produces a complete track with vocals.

Following the original ACE-Step 1.5 released in February, this new release scales the underlying DiT (Diffusion Transformer) decoder to 4B parameters. The upgrade delivers noticeably cleaner audio, better prompt adherence, and richer musicality, including the ability to handle Japanese vocals out of the box.

How it works and hardware requirements

The generation API and LoRA training code remain identical to the previous version, meaning existing workflows and setups carry right over.

Thanks to turbo distillation, the model generates complete tracks in just 8 sampling steps. On a consumer GPU, it can produce a full song in less than a minute.

However, 4B parameters require decent muscle. You will need a GPU with at least 12GB of VRAM to run it, and 20GB or more is recommended for smooth operation.

Three model variants

ACE Music released three distinct weights on Hugging Face to match different needs.

The base model, acestep-v15-xl-base, preserves the highest musical diversity. The acestep-v15-xl-sft model is fine-tuned for cleaner audio at the cost of some variety. Meanwhile, acestep-v15-xl-turbo prioritizes speed, churning out polished songs rapidly, though it is less suited for deep fine-tuning.

While internal benchmarks claim it outperforms commercial heavyweights like Udio and Suno, early hands-on testers note it feels roughly on par with Suno v4. It might not fully match Suno v5 yet, but it is closer than any open competitor before it.

Why it matters

Commercial song generators have kept high-fidelity vocal tracks locked behind monthly subscriptions, restrictive terms, and cloud-only servers. ACE-Step 1.5 XL proves that a compact 4B DiT model can now deliver competitive tracks entirely offline on everyday enthusiast hardware.

If you have a 16GB or 24GB GPU and need royalty-free music for indie games, YouTube videos, or song drafting, this is an easy install. You trade away a little cloud polish, but gain infinite generation, custom LoRA flexibility, and full privacy.