Amazing Free AI Composer: ACE-Step Now Available

A new Apache 2.0 open-source AI music generator called Ace-Tep has been released with a large 3.5B-parameter openweight model, bringing lyric generation, multi-style music, and several editing workflows into a single toolset. The headline appeal is not just that it can produce full songs quickly on high-end hardware, but that it’s designed for iteration—community members can experiment with fine-tuning and LoRA adapters, while users can repaint, edit lyrics, and extend tracks to refine results.

Performance claims set expectations clearly. On an NVIDIA A100, the system generates about a minute of audio in a little over two seconds; an RTX 4090 can do under two seconds for a minute at lower “step” settings. A MacBook M2 Max can run it too, but noticeably slower. The model’s quality knob is tied to inference steps: pushing to 60 steps roughly doubles generation time, and higher settings can increase pronunciation quirks. Even so, the creator’s testing suggests that consumer-grade hardware can still produce usable outputs—just not with one-click convenience.

Beyond raw generation, Ace-Tep is positioned as a music-production assistant. It supports mainstream music styles using multiple description formats (short tags, descriptive text, and use-case scenarios) and claims coverage across up to 19 languages, with top-performing languages including English, Chinese, Russian, Spanish, Japanese, German, French, Portuguese, Italian, and Korean. The model can generate realistic instrumental tracks with complex multi-instrument arrangements while maintaining coherence, and it also supports vocal styles.

A standout feature is audio inpainting: users can upload audio, remove a chunk, and have the model regenerate the missing section so the surrounding material stays consistent. There’s also “lyric editing,” described as a flowchart-style method that allows local lyric modifications while preserving melody and vocals—useful when generated lyrics are close but not perfect. For longer projects, the system includes repainting (regenerating a time window), editing tags, and extension workflows to build songs incrementally.

The ecosystem is also part of the pitch. Two LoRA adapters—lyrics-to-vocal and text-to-sample—are presented as demonstrations of subtask fine-tuning. Planned or upcoming adapters include a rap-focused fine-tune aimed at rap generation and “AI rap battles,” plus Stem gem (a control net) for generating individual instrument stems from multi-track data, and singing-to-accompaniment to turn raw vocals into a complementary backing track.

Hands-on tests show the tradeoffs. With mid-level settings, the model often produces a verse-and-hook structure rather than fitting every lyric line into a short target duration. It handles country ballads more smoothly than rap, where lyric delivery and pronunciation can break down. Still, the outputs can be listenable, and the workflow encourages fixing mistakes through higher steps, repainting, and targeted extensions—turning generation into an iterative creative loop rather than a single-shot result. For developers and AI music enthusiasts with substantial GPUs, Ace-Tep is framed as one of the most customizable open-source music generation releases available now, even if it doesn’t yet match the polish of top closed-source competitors.