Quick Answer
ACE-Step 1.5 is the only free AI music generator that is both genuinely capable and commercially usable. It releases model weights under MIT, runs in about 6GB of VRAM across Windows, Mac, AMD and Intel hardware, and generates full songs with vocals in over 50 languages.
ACE-Step 1.5 is the only free AI music generator that is both genuinely capable and commercially usable. It releases its model weights under MIT, runs in about 6GB of VRAM on Windows, Mac, AMD or Intel hardware, and generates full songs with vocals in over 50 languages. Its own paper shows it beats Suno on audio quality while losing on prompt adherence.
Key takeaways
- MIT licensed weights, so anything you generate is yours to sell. Most free music models are not.
- Runs in 6GB of VRAM on the turbo variant, across CUDA, AMD ROCm, Apple Silicon and Intel.
- Generates vocals with lyrics in 50+ languages, plus vocal separation and cover generation.
- Songs from 10 seconds to 10 minutes, in under 10 seconds on an RTX 3090.
- Honest weakness: it follows prompts less reliably than Suno, per its own benchmarks.
What ACE-Step actually does
ACE-Step 1.5 is a local text-to-music model. You describe the music you want, optionally supply lyrics, and it generates a complete song on your own hardware. No subscription, no credits, no upload.
Its distinguishing capability among free tools is vocals. Most open music models produce instrumental output only. ACE-Step generates sung vocals with your lyrics across more than 50 languages, and adds vocal-to-backing-track conversion, cover generation and track separation.
Two version numbers circulate and they carry different licences: ACE-Step v1-3.5B weights are Apache-2.0, and ACE-Step 1.5 weights are MIT. Both permit commercial use, but state which you tested if you are writing about it.
What you need to run it
| Configuration | VRAM |
|---|---|
| Turbo | ≤6GB |
| Turbo + 0.6B language model | 6–8GB |
| Standard / SFT | 8–16GB |
| XL variant | 20GB+ |
Model sizes are roughly 4.7GB for the 2B DiT, and about 9GB for the XL 4B in bf16.
Hardware support is unusually broad. Most local AI music tools are CUDA-only. ACE-Step runs on NVIDIA (CUDA), AMD (ROCm), Apple Silicon (MLX), Intel (XPU), and supports CPU offload for machines short on video memory.
Speed is genuinely fast: under 2 seconds per full song on an A100, and under 10 seconds on an RTX 3090. It can batch up to 8 songs at once.
Song length runs from 10 seconds to 10 minutes, which is far beyond what most free tools manage.
How to use it
ACE-Step ships as a Python project with a web interface. The broad shape:
- Clone the repository and install dependencies
- Download the model weights (start with the turbo variant at ~4.7GB)
- Launch the web UI
- Enter a style prompt, optionally add lyrics, set a duration and generate
Start with turbo. It fits 6GB, generates fastest, and is the right variant for learning whether the output suits you before committing more disk and memory to the XL model.
If you already run local AI models, the workflow will feel familiar. If you do not, our guide to running AI locally with Ollama covers the general concepts, though ACE-Step is a separate installation.
How good is it, honestly
ACE-Step's own paper benchmarks it against Suno, and it is refreshingly candid about where it loses:
| Metric | ACE-Step 1.5 | Suno v5 |
|---|---|---|
| AudioBox CU (quality) | 8.09 | n/a |
| AudioBox PQ (quality) | 8.35 | n/a |
| Style alignment | 39.1 | 46.8 |
| Lyric alignment | 26.3 | 34.2 |
Blind human A/B testing via Music Arena placed ACE-Step 1.5 between Suno v4.5 and v5.
The honest reading: ACE-Step produces excellent audio, scoring highest on both quality measures. But it follows instructions less reliably: style and lyric alignment both trail Suno meaningfully. In practice that means generating more attempts to get what you actually asked for.
For a free unlimited tool, generating five variations costs nothing but time. For a paid service metering credits, prompt adherence matters more. That tradeoff is the real difference.
Note these figures come from the developers' own paper, so treat them as a claim rather than independent verification.
Why the licence matters more than the benchmarks
ACE-Step's real advantage over the alternatives is not quality. It is that you can legally sell what it makes.
| Model | Weights licence | Sell the output? |
|---|---|---|
| ACE-Step 1.5 | MIT | Yes |
| MusicGen / AudioCraft | CC-BY-NC 4.0 | No |
| Magenta RealTime 2 | CC-BY-4.0 | Yes, with attribution |
| Stable Audio Open | Stability Community | Under $1M revenue only |
| Magenta (original) | n/a | Archived January 2026 |
MusicGen is the trap. Meta releases AudioCraft's code under MIT but its model weights under CC-BY-NC 4.0, which prohibits commercial use. It is the most-recommended free music generator and you cannot monetise its output. Our guide to free AI music you can actually sell covers this in full.
MusicGen also needs at least 16GB of VRAM for its medium model, produces mono audio, and does not generate usable vocals. ACE-Step is less demanding and more capable on every axis, on top of being commercially clean.
ACE-Step versus paying for Suno
Suno Pro is $8 a month with 2,500 credits, roughly 500 songs. Premier is $24 for 10,000 credits.
Suno wins on prompt adherence, requires no hardware, and works in a browser. Its paid tiers grant you ownership.
But note Suno's free tier does not: its help centre states that on the Basic tier "we retain ownership of the songs you generate," and rights are not retroactive when you upgrade. Generate on the paid plan from the start if you intend to release anything.
ACE-Step wins on unlimited generation, zero cost, privacy, song length up to 10 minutes, and unambiguous MIT rights with no subscription.
Suno wins on getting what you asked for first time, and requiring no GPU.
Which should you choose?
Choose ACE-Step 1.5 if you have 6GB of VRAM, want unlimited free generation, need clean commercial rights, or work with long-form music.
Choose Suno Pro at $8 if you lack suitable hardware, value prompt precision, or generate occasionally enough that $8 beats the setup effort.
Avoid MusicGen for anything commercial, regardless of how often it is recommended.
Avoid Udio if you need to release music. After Universal Music's October 2025 settlement it disabled downloads and now operates as a walled garden.
The verdict
ACE-Step 1.5 is the answer to a question that had no good answer a year ago: free, local, capable AI music you can legally sell.
It is not the best music generator available. Suno follows prompts better and needs no hardware. But it is the best free one by a wide margin, and it is the only one in that category whose licence lets you build a business on the output. For anyone generating at volume, that combination is hard to beat.
Related reading: free AI music you can actually sell and best AI music generators.
Pricing verified against each vendor's own pricing page in September 2026. Plans change often, so check the vendor's page before you buy.
Frequently Asked Questions
Is ACE-Step free for commercial use?
Yes. ACE-Step 1.5 releases its model weights under MIT, so anything you generate is yours to sell. The earlier v1-3.5B release uses Apache-2.0, which also permits commercial use. This is unusual: most free AI music models restrict commercial use through their weights licence.
What hardware does ACE-Step need?
The turbo variant runs in 6GB of VRAM or less, with standard models wanting 8 to 16GB and the XL variant 20GB+. It supports NVIDIA CUDA, AMD ROCm, Apple Silicon via MLX, Intel XPU and CPU offload, which is far broader than most local music tools.
Is ACE-Step better than Suno?
On audio quality, its own paper scores it highest on both AudioBox measures, and blind testing placed it between Suno v4.5 and v5. On prompt adherence it loses: style alignment 39.1 against Suno's 46.8, and lyric alignment 26.3 against 34.2. It sounds excellent but follows instructions less reliably.
Can ACE-Step generate vocals?
Yes, with lyrics in more than 50 languages, plus vocal-to-backing-track conversion, cover generation and track separation. This sets it apart from most free models including MusicGen, which does not produce usable vocals.
How long can ACE-Step songs be?
From 10 seconds to 10 minutes, generated in under 10 seconds on an RTX 3090. It can also batch up to 8 songs simultaneously, which is well beyond what most free tools manage.