ACE-Step
An open-source music generation foundation model combining diffusion with a lightweight linear transformer, released May 2025 by ACE Studio and StepFun.
- Up to
- 240 seconds
- Weights
- Downloadable
- API
- No
- Commercial use
- Vendor states commercial use is permitted
It generates up to about four minutes of music with vocals across 19 languages in roughly 20 seconds on an A100 (27x real time), and supports lyric editing that preserves melody and accompaniment, section repainting, a Lyric2Vocal LoRA for vocal-only stems, and voice cloning. Apache-2.0 licensed, making it the most permissive capable open music model.
Apache License 2.0, which permits commercial use, modification and redistribution with no revenue cap and no attribution-in-product requirement beyond the licence notice.
Specification
- What it takes in
- text, lyrics, audio
- What it makes
- audio
- Maximum duration
- 240 seconds
- Controls
- lyrics-input, style-prompting, instrumental-only, inpainting-remix, reference-audio, extend
- Weights
- Downloadable
- API
- No
- Pricing
- Free; self-hosted. GPU cost only.
- Released or updated
- 2025-05
Strengths
- Apache-2.0: unrestricted commercial use of the model
- Very fast - about 20 seconds for 4 minutes of audio on an A100
- Vocals in 19 languages
- Lyric editing and section repainting without regenerating the whole track
- Lyric2Vocal LoRA produces vocal-only stems
Weaknesses
- Training data provenance is not documented, so output clearance risk is on you
- Requires your own GPU and setup
- No hosted first-party API or support
- Sample rate and output format not documented in the README
- Quality below Suno v5.5 and Lyria 3 Pro
Best for
- Teams needing commercially licensed open weights they can self-host
- High-volume or offline generation
- Fine-tuning to a house style
Where the sources disagree
Recorded rather than resolved. Picking the tidier number would hide the fact that the vendor's own pages do not agree.
- The README states 4 minutes is the default configurable maximum rather than a hard cap, so 240 seconds should be read as the documented default ceiling.
What to understand before using it
Explainers naming the mechanisms this model's own documentation mentions.
MCP servers that front ACE Studio / StepFun
Servers in the media MCP directory whose own description or homepage names this vendor. Absence of a server here is not evidence that none exists: most publish no tool list at all.
-
Audiolla
Self-hosted MCP server: audio stems, mastering, MIR analysis, DSP, MIDI, speech tools.
-
Replicate
Replicate MCP server - image, video, audio, speech, LLM, vision, upscale, 3D, training. 36 tools.
-
avots
Image, video, audio, face-swap, talking avatars and chat across 300+ AI models, one balance.
-
boostaudio
AI audio tools for music producers, stem splitting, vocal removal, BPM & key detection, audio-to-MIDI, format conversion, trimming, video-to-audio extraction and AI song
See it from the vendor
Shotcraft hosts no sample output. Judging a model means watching its own showreel with your own eyes, and a clip re-hosted here would be a clip we chose.
Nearby in the directory
AIVA
Generates instrumental music from style presets and influence uploads for video, game and content creators, exporting MP3 and MIDI on lower tiers and
Music and sound
Beatoven.ai maestro
Generates royalty-free background music and text-prompted sound effects for video, podcast, game and advertising work, downloadable as MP3 or WAV with
Music and sound
Eleven Music (music_v2)
Generates songs with vocals or instrumentals from a natural-language prompt, from user lyrics, or from a structured JSON 'composition plan' that speci
Music and sound