Introduction

Feel that shift in the pocket? StepFun just put the music model front and center. Together with ACE Studio, they’re presenting StepAudio 3 Music—StepFun’s first music-generation foundation model—built to turn a prompt and lyrics into a complete song, not a teaser loop.

The hook is familiar if you’ve been watching the open lane: plan first, then perform. Their ABC-COT path sketches structure and arrangement before synthesis, so genre, vocals, instruments, key, and BPM aren’t afterthoughts bolted onto a lucky take.

What’s New in StepAudio 3 Music

StepAudio 3 Music sits inside the broader StepAudio 3 family, but this drop is the music-first story: MoE architecture with an AR + DiT generation path, plus ABC-COT so natural-language creative intent becomes an editable music plan before audio lands.

Four workflows in one model:

  • Song generation — prompt + lyrics → full track with vocals and accompaniment
  • Instrumental generation — arrangement without a lead vocal
  • Music cover — reimagine an existing song
  • Vocal-to-song arrangement — dry vocals → full production around them

You can describe genre, mood, vocal character, instruments, key, BPM, and song structure in plain language. Generate a version, rewrite the prompt, edit the ABC notation, and iterate toward the sound in your head. StepFun notes an ABC-COT API is coming soon. Output target on the product page: 48 kHz stereo full songs you can audition and download—not fragments.

Their music page and technical report lean hard on controllability plus quality: SOTA claims on Audiobox (quality) and MuQ-Similarity (controllability), competitive SongBench, and on a preliminary Artificial Analysis Music Arena Vocals board a Quality Elo of 1105—behind only Suno V5.5 and Mureka in their published comparison, ahead of Suno V5 and several MiniMax-class systems. Tracks can run up to about 5 minutes 30 seconds in the report’s training/SFT framing.

Pro Tip

Treat ABC-COT like a session sketch: nail structure and arrangement intent in the plan, then bounce. When the arrangement fights you, edit the ABC or rewrite the prompt before burning another full render—same muscle memory as fixing a chart before the vocal take.

How to Get Started with StepAudio 3 Music

  1. Open the interactive demo and listen to planned-then-sung examples: StepAudio 3 Music demo
  2. Try a short brief: genre + mood + vocal character + BPM + a verse/chorus lyric stub.
  3. Iterate—change the prompt, then lean on ABC edits when the API lands for finer arrangement control.
  4. Dig the technical report for tokenizer/DiT/ABC-CoT details if you’re integrating via StepFun’s open platform.

Demo: static.stepfun.com/blog/stepaudio3/music · Announcement:

· Tech report: arXiv:2609.16034 · Platform: stepfun.com

Conclusion

StepAudio 3 Music is the StepFun × ACE Studio music drop that finally talks like a producer: structure first, synthesis second, four workflows that cover song, instrumental, cover, and vocal-to-arrangement. ABC-COT is the deep low-end under the claim—editable planning before the bounce.

Open the demo, drop a prompt that actually means something, and let that unfiltered motion hit your monitors.

—Anabel ♡