Create music in the official ACE-Step 1.5 local Gradio interface, listen to the result and save its WAV and settings. The first example generates a 30-second instrumental with the publisher turbo model and the optional language model disabled.
First track: 30 seconds, one instrumental, 8 steps, shift 3, seed 42. No optional 5Hz language model.
| Machine | ACE-Step v1.5 turbo |
|---|---|
| NVIDIA · 4–6 GB VRAM + 32 GB RAM | DiT-only; automatic CPU offload/quantization. Start with one 30-second clip. |
| NVIDIA · 8–12 GB VRAM + 32 GB RAM | Recommended first setup; keep auto memory controls and LM disabled. |
| AMD GPU PC · 32 GB system RAM | Native CPU fallback. Documented ROCm installation is outside this route. |
| Intel GPU PC · 32 GB system RAM | Native CPU fallback. This guide does not configure Intel XPU. |
| Apple Silicon · 16 GB unified memory | Estimated starting capacity; 30 seconds, batch 1, LM disabled. |
| Apple Silicon · 24–32 GB unified memory | Recommended memory allowance for native Apple inference and OS overhead. |
| CPU-only PC · 16 / 32 GB system RAM | 16 GB starting estimate; 32 GB recommended. Expect substantially longer waits. |
The publisher lists 4 GB DiT-only GPU tiers with offload. Mac and CPU figures are planning estimates, not measured peaks. About 6.34 GB of selected components (up to 10.1 GB in the default bundle) plus the environment require disk headroom; reserve 35 GB.
Describe music in the official ACE-Step 1.5 local browser interface, generate a result, then listen and download the WAV. The first example is a 30-second instrumental using the original turbo package with the optional language model off. Native Windows, Apple Silicon and Ubuntu paths include explicit CPU fallbacks for AMD/Intel PCs.
This setup uses the package documented for this task. Review its source and supported platforms before starting.
Save your machine in My Hardware to get an automatic starting choice. You can always choose any package yourself.
30 seconds, batch 1, 8 turbo steps, LM off. 4 GB NVIDIA requires automatic offload/quantization plus at least 16 GB host RAM; 8 GB VRAM + 32 GB RAM recommended. Mac/CPU capacities are estimates accounting for models, activations and OS. Longer music uses more memory.
Choose your computer first, then its GPU path. The instructions below update to match your selection.
Native CPU fallback on AMD/Intel PCs; no GPU acceleration claimed. First workload: 30 seconds, one instrumental, 8 steps, LM off.
Use 64-bit Windows 11. Install uv with the linked official instructions; run this command in PowerShell, then reopen PowerShell so PATH refreshes. Install Git for Windows with its default command-line integration. This route runs on the CPU; AMD and Intel graphics do not require a GPU runtime.
uv installation Git for Windows installerwinget install --id=astral-sh.uv -e
winget install --id=Git.Git -eIn PowerShell, create the project and clone the official repository. Keep the generated source-revision.txt and uv.lock with the project so you can identify this installation later. This guide selects the original acestep-v15-turbo; do not substitute the XL package. Leave at least 35 GB free for the app, model components, Python packages and download cache.
ACE-Step source and installationNew-Item -ItemType Directory -Force "$HOME\yourunai-music" | Out-Null
Set-Location "$HOME\yourunai-music"
git clone --depth 1 https://github.com/ACE-Step/ACE-Step-1.5.git
Set-Location ACE-Step-1.5
git rev-parse HEAD > source-revision.txtIn your browser, download ffmpeg-release-essentials.zip from the linked Gyan builds page, which FFmpeg lists as a Windows binary distributor. Extract the ZIP with File Explorer. First create the nested folders tools, ffmpeg and bin inside this project using File Explorer. Open the extracted inner folder, then bin; copy ffmpeg.exe, ffprobe.exe and ffplay.exe into that project tools\ffmpeg\bin folder. In the same PowerShell window run the commands below. This changes PATH for this window only.
FFmpeg official download choices Gyan FFmpeg Windows builds$env:Path = "$PWD\tools\ffmpeg\bin;$env:Path"In PowerShell from the ACE-Step-1.5 folder, install stable Python 3.11 and sync the official dependencies. On Windows/Linux these include CUDA-capable PyTorch wheels even for the CPU route; CPU execution does not require an NVIDIA card. On Apple Silicon the project installs its macOS dependencies. Keep the resulting environment for subsequent runs; do not replace PyTorch with unrelated ROCm or XPU wheels in this environment.
Official package dependencies uv installationuv python install 3.11
uv sync --python 3.11In PowerShell, write this .env in the new ACE-Step-1.5 folder, then start the application. It disables the optional 5Hz language model for this first DiT-only workload and binds the interface to this computer. Leave this window running. In your browser open http://127.0.0.1:7860 when the terminal reports it is ready. Keep the same folder and environment on restarts.
ACE-Step environment and launch options@'
ACESTEP_INIT_LLM=false
ACESTEP_CONFIG_PATH=acestep-v15-turbo
ACESTEP_DOWNLOAD_SOURCE=huggingface
SERVER_NAME=127.0.0.1
PORT=7860
LANGUAGE=en
'@ | Set-Content -Encoding utf8 .env
uv run --no-sync acestep --server-name 127.0.0.1 --port 7860 --language en --config_path acestep-v15-turbo --init_llm false --batch_size 1 --backend ptIn the browser’s Service Configuration, select Main Model Path acestep-v15-turbo and Device cpu. Leave Initialize 5Hz LM unchecked. Click Initialize Service and wait for the green ready status. The app manages checkpoints/acestep-v15-turbo (4.79 GB DiT), checkpoints/Qwen3-Embedding-0.6B (1.21 GB encoder) and checkpoints/vae (337 MB audio VAE). It may download the full 10.1 GB default bundle including the unused LM; disabling the LM avoids loading it, not necessarily downloading it. CPU inference is supported but can take substantially longer; use 32 GB RAM when available. Do not proceed while initialization reports a missing file or download error.
Publisher weights and component folders Hardware tiers and offload settingsIn the music tab select Custom mode. Enter this Caption: “A calm acoustic instrumental with fingerpicked guitar, soft piano and gentle brushed drums, warm evening mood, no singing.” Leave Lyrics empty, enable Instrumental, and use Vocal Language unknown. Set Audio Duration to 30 seconds, Batch Size 1 and BPM 90. Turn Think, CaptionRewrite, CoT Metas, CoT Language and AutoGen off. Avoid Create Sample and Format for this LM-disabled setup. These are browser settings, not terminal commands.
Official Gradio generation controlsIn Advanced Settings set Inference Steps 8, Shift 3, Inference Method ode and Audio Format wav. Uncheck Random Seed and enter Seed 42. Click Generate Music once. Wait for the finished audio player and generation status; leave the terminal running. The duration is a target, so inspect the actual file rather than treating the setting as a measurement.
Gradio advanced settingsPress Play in the result’s audio player. Listen through the beginning and end; compare the audible instruments and mood with your caption, check that vocals are absent, and inspect its duration. Use the player’s Download control to save the WAV. In your browser’s download/save dialog save or move that file into $HOME\yourunai-music\ACE-Step-1.5\exports\music.wav; create exports if needed. Click Save for the result’s metadata JSON and keep that download beside the WAV as music-settings.json. Open the downloaded music.wav in your normal audio player and play it again; a browser waveform alone is not the exported result.
Within the same session use Apply These Settings to UI on a result to recover its controls, change one detail in the Caption or Seed, and click Generate Music again. Save each chosen result under a different filename. To stop, press Ctrl+C in the original terminal and wait for the shell prompt. Tomorrow open PowerShell and run the commands below, then initialize the model and re-enter the saved prompt/settings from music-settings.json in Custom mode. Fixed seeds help comparison, but different hardware or app versions can still change the output.
Set-Location "$HOME\yourunai-music\ACE-Step-1.5"
$env:Path = "$PWD\tools\ffmpeg\bin;$env:Path"
uv run --no-sync acestep --server-name 127.0.0.1 --port 7860 --language en --config_path acestep-v15-turbo --init_llm false --batch_size 1 --backend ptexports/music.wav plays outside the browser, and you have compared its instruments, lack of vocals and duration with the request. music-settings.json records the result’s settings. The local interface can be stopped with Ctrl+C and restarted from the same project.
Sources checked September 28, 2026. These are editorial review drafts, not physical hardware tests. Keep the downloaded portable archive for Windows NVIDIA or Mac; source installations retain source-revision.txt and uv.lock. No generated audio or timing is claimed.
Keep the original terminal open and read its startup error. Wait for the local URL before opening http://127.0.0.1:7860. If port 7860 is already occupied, stop the other local service or use --port 7861 in the launch command and open that port.
Keep Internet access available and retry the initialization after checking disk space. The selected package is acestep-v15-turbo in the official ACE-Step/Ace-Step1.5 repository; do not replace it with an XL checkpoint. Wait for green status before generating.
Return to batch 1 and 30 seconds, leave Initialize 5Hz LM off, and close other memory-heavy applications. On NVIDIA retain the automatic CPU offload/quantization settings for the detected tier. A 4 GB GPU still needs host RAM; prefer 32 GB. On Apple use platform defaults and reduce the trial duration to 10 seconds if needed.
Use Custom mode, disable Think, CaptionRewrite, CoT Metas and CoT Language, and skip Create Sample/Format. Empty Lyrics, enable Instrumental and keep the no-singing instruction in the Caption. Generate a new seed and listen again; the checkbox is not an audio-quality guarantee.
CPU inference is supported but does not gain CUDA speed from the installed wheel. Keep the 30-second, batch-1 trial and allow it to finish. The official environment includes CUDA packages on Windows/Linux; a smaller custom dependency build is outside this guide.
Original instructions, model files, and compatibility notes behind this setup.
Source-backed package and platform-route plan for private local chat with Qwen3.5 · 9B. Preserves the official Ollama 9B package, a smaller 4B family fallback, and a clearly labeled community uncensored GGUF option.
Run Qwen3.8 27B with Ollama, connect it to Cline in VS Code, and make one reviewable code change in a tiny local project.
Generate images from text prompts in ComfyUI with the fast distilled or more adjustable base 4B model. Save your results and reuse the visual workflow.