AI Potluck
Back to Gap Map Model components / Image, video, 3D & music generation

ACE-Step

ACE-Step
open weights / Overall score: 1.8

ACE-Step is an open music generation model family from the ACE Studio and StepFun collaboration that writes full songs, vocals included, from lyrics and a style prompt. ACE-Step 1.5 pairs a diffusion transformer, up to 4B parameters in its XL variants, with a language model that plans the song, and supports cover generation, repainting, completion and LoRA training.

Openness

3 high confidence
3.0
weights
open(the ACE-Step 1.5 base, SFT, turbo and XL checkpoints, the planning language models and ACE-Step v1 download from Hugging Face without a gate)
data
described(the 1.5 card describes licensed, royalty-free and synthetic training music
code
partial(ACE-Step 1.5 publishes LoRA training only
license
MIT(ACE-Step 1.5 code and weights)+Apache-2.0(ACE-Step v1 code and weights)

ACE-Step's weights and code are released under MIT and Apache 2.0, and the generated music may be used commercially. The training data is described but not released, and the current release publishes only LoRA training rather than its full pipeline.

Adoption

3 high confidence
3.0

Adoption is measured as Hugging Face downloads across the ACE-Step 1.5 and v1 checkpoints, led by the main 1.5 repository. The LoRA adapters are left out.

Capability

1 medium confidence
1.0

ACE-Step generates complete songs with vocals, but it has no placement on the public music arenas, where MusicGen at least appears. That leaves its quality against the closed services unmeasured.

Verified 2026-09-26