forkalt.

MITaudioRuns locallyLM + DiT (2B DiT default; 4B XL variants)

ACE-Step 1.5: Open Music Model Specs & License

Open MIT music foundation model for local song generation on consumer GPUs.

ACE-Step 1.5 documentation preview image from the official repository
Official project image. Source · License: MIT.
License
MIT
Parameters
LM + DiT (2B DiT default; 4B XL variants)
HF likes
907 (2026-10-18)
First seen
2025-09-04

TL;DR. ACE-Step 1.5 is a audio model from ACE-Step (ACE Studio / StepFun). Parameters: LM + DiT (2B DiT default; 4B XL variants).License: MIT. It is listed under Suno.

DeveloperACE-Step (ACE Studio / StepFun)
Typeaudio
ArchitectureHybrid language-model planner plus Diffusion Transformer decoder
ParametersLM + DiT (2B DiT default; 4B XL variants)
Modalitiesaudio, music, text
LicenseMIT
Hardware<4 GB (vendor README for the default local path); XL ≥12 GB with offload — README (checked 2026-10-18) says the model runs locally with less than 4 GB VRAM, and XL (4B DiT) needs ≥12 GB with offload or ≥20 GB without. CUDA is recommended; MPS, ROCm, Intel XPU, and CPU are also listed. forkalt has not measured VRAM. (not independently tested)
Hugging Face likes907 (as of 2026-10-18)
Model cardhttps://huggingface.co/ACE-Step/Ace-Step1.5
First seen2025-09-04

ACE-Step 1.5 is an open music generation model. Official hosts we opened on 2026-10-18: ace-step/ACE-Step-1.5, ungated card ACE-Step/Ace-Step1.5, and acemusic.ai. It hangs off Suno alternatives. That mapping is editorial. We have no sourced quality comparison against Suno.

Key facts (2026-10-18)

Field Value Source
GitHub stars / forks 13,728 / 1,807 GitHub API ace-step/ACE-Step-1.5
Created / last push 2025-09-04T02:00:07Z / 2026-10-11T14:11:22Z same
License file MIT LICENSE
HF likes / downloads 907 / 75,886 HF API ACE-Step/Ace-Step1.5
HF gated false same
HF card license mit cardData.license
Pipeline tag text-to-audio HF API
Latest GitHub tag v0.1.8, 2026-05-18T14:04:34Z GitHub Releases

The GitHub repo is newer than the May 2026 tag (last push October 2026). Prefer the README on main over the tag date.

What the README claims (vendor)

Copied, not independently tested: hybrid LM planner + DiT decoder; songs from short loops to 10 minutes; 50+ lyric languages; cover, repaint, vocal-to-BGM, LoRA from a few songs; default local path <4 GB VRAM; XL 4B DiT ≥12 GB with offload. Quality “between Suno v4.5 and Suno v5” stays vendor self-reported.

Hosted acemusic.ai is a different product from downloading weights.

Install (from the official README, not yet tested)

git clone https://github.com/ACE-Step/ACE-Step-1.5.git
cd ACE-Step-1.5
uv sync

Python 3.11–3.12. CUDA recommended. forkalt has not run this.

Deploy options

FAQ

Are ACE-Step 1.5 weights public?

Yes. Hugging Face ACE-Step/Ace-Step1.5 is ungated (created 2026-01-23T09:59:48Z; 907 likes and 75,886 downloads in the 2026-10-18 API check; card license mit; pipeline_tag text-to-audio). Code is github.com/ace-step/ACE-Step-1.5 (MIT LICENSE file).

Can I use it commercially?

The GitHub LICENSE is MIT (checked 2026-10-18). Hugging Face cardData.license is mit. The README still warns about stylistic similarity and misuse. Not legal advice.

Does it beat Suno?

We do not claim that. The README says quality is "between Suno v4.5 and Suno v5" on the authors' metrics. That is vendor self-reported. There is no sourced bake-off against Suno on this page.

What hardware does the README require?

Default local path is quoted as under 4 GB VRAM. XL 4B DiT variants need ≥12 GB with offload. Python 3.11–3.12. CUDA recommended; MPS / ROCm / Intel XPU / CPU are listed. Speed figures (under 2 s/song on A100, under 10 s on RTX 3090) are vendor claims. forkalt has not run them.

Related pages

Sources & verification

Last verified 2026-10-18 · Last updated 2026-10-18· Edited by forkalt editors · Editorial policy

License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.