Four free AI music posts have gone up on this site in August, and the model that working producers keep quietly recommending was not one of them. ACE-Step 1.5 is an MIT-licensed music foundation model that turns lyrics and a few comma-separated tags into a finished song, runs on a 4GB graphics card, and ships with a Gradio web UI, a REST API, and a VST3 plugin you can load directly in a DAW.
The reason it deserves your afternoon is the license. Most of the free music models released this year arrive with strings attached, whether that is an attribution clause, a revenue ceiling, or a non-commercial rider. ACE-Step 1.5 is plain MIT on both the code and the weights. If you render a track for a client video today, nothing in the license stops you shipping it. There is a free hosted demo you can open in a browser right now with no signup, and weights you can download and keep.
At a Glance
- What it is: an open-source text-to-music foundation model from ACE Studio and StepFun that generates 10 seconds to 10 minutes of audio from lyrics plus style tags, in 50+ languages.
- What is free: everything. Code, weights, the Gradio UI, the REST API server, the LoRA trainer, and a public Hugging Face demo Space. No account required for the Space.
- License: MIT, on the repo and the model cards. Commercial use is allowed with no attribution requirement and no revenue threshold.
- Hardware: the 2B turbo decoder is a 4.79 GB download and the project targets 6GB VRAM or less with INT8 quantisation and CPU offload. The higher-quality 4B XL decoder wants 12GB minimum, 20GB comfortably.
- The catch: the “better than most commercial models” claim is the team’s own, from their own evaluation. No independent leaderboard has confirmed it.
What ACE-Step 1.5 actually does
The architecture is a two-stage hybrid. A small language model built on Qwen3 acts as a planner: it takes a short user query and expands it into a full song blueprint, writing metadata, lyrics and captions through a chain-of-thought pass. That blueprint then conditions a Diffusion Transformer that renders the actual audio. Three planner sizes ship, at 0.6B, 1.7B and 4B parameters, and you pick one based on how much VRAM you have to spare.
Straight text-to-music is only part of it. The same stack does cover generation from a reference track, repaint (regenerate one section of an existing song while keeping the rest), stem separation, multi-track layering, vocal-to-backing-track conversion, LRC lyric timestamping, and audio understanding that extracts BPM, key and time signature from a file you feed it. That editing surface is what separates it from the one-shot generators covered here recently, such as MiniMax Music 3.
What it needs to run
The project publishes a tier table rather than a single number, and the Gradio UI auto-selects a configuration for your card. In plain terms:
- 6GB or less: 2B turbo decoder only, planner disabled, INT8 quantisation plus full CPU offload.
- 6 to 8GB: 2B turbo plus the 0.6B planner on the PyTorch backend.
- 8 to 16GB: 2B turbo or SFT, with the 0.6B or 1.7B planner on vLLM.
- 16 to 24GB: the 4B XL decoder becomes usable, with CPU offload below 20GB.
- 24GB and up: XL SFT plus the 4B planner, everything resident, best quality.
Beyond NVIDIA, the repo supports Apple Silicon through MLX, AMD via ROCm, Intel Arc through XPU, and plain CPU. Portable packages with dependencies pre-installed exist for both Windows and macOS, which matters if you have ever lost an evening to a PyTorch version conflict.
How it compares to the other free options
| Model | Free? | License | Entry VRAM | Max length |
|---|---|---|---|---|
| ACE-Step 1.5 (2B turbo) | Weights + demo, yes | MIT | ~6GB (less with offload) | 10 minutes |
| MiniMax Music 3 | Weights, yes | Custom: attribution required, authorisation needed above $20M revenue | ~8GB with offload | 5 minutes |
| Stable Audio 3 Small | Weights, yes | Stability community license | CPU-capable | Short clips |
| Suno | No, paid tiers | Proprietary, cloud only | None (hosted) | Varies by plan |
If you only want a browser tab and a result, the HappyShrimp write-up is still the fastest route. If your machine has no GPU at all, Stable Audio 3 Small remains the CPU pick. ACE-Step 1.5 is the one to install when you want to own the pipeline.
Try ACE-Step 1.5 free today
Zero install, 30 seconds: open the ACE-Step v1.5 Playground on Hugging Face. It runs on ZeroGPU, defaults to the XL turbo model, and gives you Simple, Custom, Cover and Repaint modes with prompt, lyrics and optional reference-audio fields. No login needed.
Local install: you need Python 3.11 or 3.12. Install uv, then run four lines. Models download automatically on first launch.
curl -LsSf https://astral.sh/uv/install.sh | sh
git clone https://github.com/ace-step/ACE-Step-1.5.git
cd ACE-Step-1.5 && uv sync
uv run acestepThe Gradio UI opens at http://localhost:7860. Swap the last line for uv run acestep-api to get the REST server on port 8001 instead. Windows users can run start_gradio_ui.bat; on Apple Silicon use start_gradio_ui_macos.sh, which routes through MLX.
The rest of the links you will want: the GitHub repository (12.4k stars, 77 contributors), the main weights repo and the 4B XL turbo decoder on Hugging Face, the full installation guide covering ROCm and Intel GPUs, the usage tutorial, the LoRA training guide, the VST3 plugin, the technical report, and Awesome ACE-Step for community ComfyUI nodes and alternative UIs.
The honest trade-offs
- The quality claim is unverified. The README places output “between Suno v4.5 and Suno v5”. That is the team’s own measurement on their own benchmark. Treat it as a hypothesis and A/B it against your own reference tracks before you rely on it.
- The 4GB headline is the small model. Six VRAM tiers exist, and the sub-6GB configuration disables the planner entirely and runs INT8 with full CPU offload. The output that gets compared to commercial models is the 4B XL decoder, which needs a 12GB card at minimum.
- Release cadence has slowed. The latest tagged release is v0.1.8 from 18 May 2026, though the main branch received fixes as recently as 16 August 2026. Install from main, not the tag, if you hit a bug.
- The hosted option has a login wall. The team promotes acemusic.ai as “100% free”, and it is, but the landing page is a login screen. The Hugging Face Space is the genuinely frictionless path.
- MIT is not a copyright shield. The license clears the software. It does not clear you if you prompt for a named artist’s style and release the result commercially.
Five things to do differently
- Test your tag vocabulary on the Space in Simple mode before committing to a 4.79 GB download. Ten minutes there tells you whether the model understands the genre you actually work in.
- Write tags comma-separated with no spaces, exactly as the docs specify:
piano,happy,wedding,synthesizer,romantic. Spaces degrade parsing. - Use section tags in your lyrics.
[Intro],[Verse],[Prechorus],[Chorus],[Bridge],[Outro]drive structure far harder than prose descriptions do. - Reach for Repaint before you regenerate. If the second verse is wrong, repaint that window rather than rolling a fresh song and losing the chorus you liked.
- Train a LoRA once you have a sound. The docs quote roughly one hour on a 12GB RTX 3090 from eight reference songs, with annotation and training handled in the Gradio UI.
FAQ
Is ACE-Step 1.5 free?
Yes. The code, the model weights, the Gradio interface, the REST API and the LoRA trainer are all free under the MIT license, and the Hugging Face demo Space costs nothing and requires no account.
Can I use ACE-Step 1.5 commercially?
The MIT license permits commercial use with no attribution requirement and no revenue cap. You remain responsible for the content you generate, including any stylistic similarity to protected work.
What are ACE-Step 1.5’s system requirements?
Python 3.11 or 3.12, and a GPU with 6GB of VRAM for the 2B turbo decoder (a 4.79 GB download) or 12GB minimum for the 4B XL decoder. CPU-only, AMD ROCm, Intel XPU and Apple Silicon paths are all supported.
Does ACE-Step 1.5 run on a Mac?
Yes. There is a macOS portable package and an MLX-backed launch script for Apple Silicon.
How long can ACE-Step 1.5 songs be?
Ten seconds to ten minutes, and it can generate up to eight songs in a single batch.
For a look at how tools like this fit into a real production pipeline, this walkthrough of turning old client content into a full AI podcast and social plan covers the workflow end to end, and there is more on the Creative AI Show channel.
Accurate as of August 27, 2026 — model licences, VRAM requirements and leaderboard positions move fast. Check the repository before you ship.
No responses yet