How to run Stable Audio Open on a cloud GPU

Rent a GPU on Deploy, then open a terminal in JupyterLab or over SSH. Install the official tools repository, accept the model terms on the page that README links, and run either the Gradio UI that README prints or the Python example on the model card. Copy the audio file off the instance before you destroy it. Destroying the instance deletes its disk.

The tools README and the model card publish no GPU GB figure. The model card says Stable Audio Open 1.0 generates variable-length (up to 47s) stereo audio at 44.1kHz. Those are the model card’s figures. This page did not re-measure them.

Other music and speech installs: How to run ACE-Step 1.5 on a cloud GPU, How to run YuE music generation on a cloud GPU, and How to run Fish Speech on a cloud GPU.

The commands below are copied from the upstream files named in each step. They were not run on a CloudGPU instance for this page. Anything this page adds around them is marked. The README’s Gradio example and the model card’s Python example both pass a Hub repository name. This page does not paste that Hub id. Copy the argument from the file you are following.

What the README and the model card say about memory

The stable-audio-tools README says the repository requires PyTorch 2.5 or later, and that development is done in Python 3.10. It does not print a GPU GB figure. The model card read the same day does not print one either. pyproject.toml requires Python >=3.10,<3.11, pins torch==2.7.1 and torchaudio==2.7.1, and on Linux x86_64 takes those two packages from the cu126 index. .python-version is 3.10. This page does not treat the README’s “2.5 or later” sentence and the 2.7.1 pin as the same claim.

GET /v1/public/compute-capabilities does not list GPUs. It returns feature flags. The memory column is vram_gb from gpu-prices.json, updated_at 2026-11-12T22:09:18.532Z. GET /v1/public/gpu-catalog the same day (generated_at 2026-11-12T22:09:19.501Z) listed the same names and the same vram_gb values, plus Ascend 910B at 64, which is not in gpu-prices.json and is not listed here. The catalog response also includes an availability field. This page does not copy it. There is no upstream GPU GB cutoff to match these values against. This page does not mark a row as enough. Which listing you can start is on Deploy. The catalog had no Stable Audio template.

Catalog name vram_gb
RTX 2080 8
RTX 3080 10
RTX 2080 Ti 11
RTX 3080 Ti 12
V100 16G 16
A30 24
RTX 3090 24
RTX 4090 24
RTX 4090D 24
RTX 4080 SUPER 32G 32
RTX 5090 32
RTX 5090D 32
V100 32G 32
A100 40G 40
A800 40G 40
A40 48
RTX 4090 48G 48
RTX 4090D 48G 48
RTX A6000 48
NVIDIA RTX PRO 5000 72G 72
RTX PRO 5000 72G 72
A100 80G 80
A800 80G 80
H800 80
RTX 6000D 83
H20 96G 96
RTX PRO 6000 96G 96
RTX PRO 6000 Max-Q 96
H20-3e 141G 141

License

The tools repository LICENSE is the MIT License. Its copyright line is “Copyright © 2023 Stability AI”. The grant says “free of charge”. That sentence is the code license. It is not a CloudGPU price and not a storage allowance. The commits API on 2026-11-12 listed 3241adba4fc2a85cf5b29d9eb68d42f40a28e820 on the default branch, committed 2026-05-26T23:34:41Z. The GitHub license field the same day was MIT.

The weights are a different file. A model API read on 2026-11-12 reported gated: auto, cardData.license other, cardData.license_name stable-audio-community, and license_link LICENSE. This page does not paste the Hub id. LICENSE.md fetched from that repository the same day is titled STABILITY AI COMMUNITY LICENSE AGREEMENT, last updated July 5, 2024. The licensor’s name appears because it is the text of that file.

The introduction says the agreement is intended to allow research, non-commercial, and limited commercial uses of the models free of charge, and that it preserves free access for people or organizations generating annual revenue of less than US $1,000,000 (or local currency equivalent). Those “free” sentences and that dollar figure are the license text. They are not a CloudGPU price and not a storage allowance. Section II grants research and non-commercial use as that section defines those purposes. Section III grants commercial use and says you must register at the URL written in that section (https://stability.ai/community-license). It says the license terminates if you or your affiliates generate more than USD $1,000,000 in annual revenue, and it points at https://stability.ai/enterprise for a further license. That second dollar figure is also the license text, not a CloudGPU price. Section IV(a) requires a copy of the agreement, the attribution notice written in that section, and a “Powered by” display that names the licensor. Section IV©(iii) says you own outputs to the extent permitted by applicable law. Read the file. This page does not restate the rest of it.

Install

On Deploy, choose Raw GPU. The README says to install uv with pip install uv or from the uv docs, then clone and sync. The “Everything” line is the one that includes the Gradio extra. Copied, with the README’s $ prompt removed:

git clone https://github.com/Stability-AI/stable-audio-tools.git
cd stable-audio-tools
uv sync --extra train --extra ui

The README’s inference-only line is uv sync, and its training line is uv sync --extra train. The pip alternative it prints is pip install "stable-audio-tools[train]", which does not include the ui extra. The ui extra in pyproject.toml is gradio==6.3.0. The README says to install Flash Attention afterwards by following that project’s instructions. How to install FlashAttention on a cloud GPU is a separate page. It is not a line from this README.

The README says to accept the terms on the model page before the Gradio example. The model API reported gated: auto. How to download a Hugging Face model is the token reference. The tools README does not print a login command.

Gradio UI and SSH

The README prints two Gradio examples. One is uv run python run_gradio.py --pretrained-name followed by a Hub repository name. The other is python3 ./run_gradio.py --pretrained-name with the same name. Copy that argument from the README. This page does not paste it. run_gradio.py also accepts --model-config, --ckpt-path, --share, --username, --password, and --model-half. The --share help text is “Create a publicly shareable link”. The --model-half argument is store_true with default=True. This page did not run the script and does not pass --share.

run_gradio.py calls interface.launch with share and auth and does not pass server_name or server_port. Gradio 6.3.0, the version the ui extra pins, documents in Blocks.launch that a None server name uses 127.0.0.1 and a None server port searches for an available port starting at 7860. This page did not run the launch and does not claim the search stays on 7860. Read the URL the process prints.

To keep that process after the SSH session closes, this page wraps the README’s uv run form in tmux. The README does not print tmux, and the Hub id is still for you to copy from the README:

tmux new -s audio
cd stable-audio-tools
uv run python run_gradio.py --pretrained-name <name-from-the-readme>

The tunnel command is not in the tools README. It is the form in how to use SSH port forwarding on a cloud GPU, using the port the process prints:

ssh -N -L <port>:127.0.0.1:<port> -p <ssh-port> <ssh-user>@<ssh-host>

Then open http://127.0.0.1:<port> on your computer.

Python inference

The model card’s “Using with stable-audio-tools” section prints a script that calls get_pretrained_model with a Hub repository name, then generate_diffusion_cond, then torchaudio.save("output.wav", output, sample_rate). Copy the Hub argument from the model card. This page does not paste it and did not run the script. The numeric arguments in that block, including steps and the duration fields, are the model card’s example. This page does not restate them as a measured result.

The same card’s diffusers section prints pip install -U diffusers and a script that builds StableAudioPipeline.from_pretrained with a Hub repository name, then sf.write("hammer.wav", ...). Copy the Hub argument from the model card. This page does not paste it.

The README’s training section is a separate path. It says the training code requires a Weights & Biases login and prints wandb login, train.py, and --checkpoint-every with a default of 10000. This page does not start a training run.

ComfyUI route

The ComfyUI audio example has a section headed “Stable Audio Open 1.0”. That section has two download links. This page does not paste either Hub path. Copy them from the example. The example says to save the first download as t5_base.safetensors in ComfyUI/models/text_encoders/, and the second as stable_audio_open_1.0.safetensors in ComfyUI/models/checkpoints/. The first link is a different repository from the second. The workflow is in stable_audio_example.flac on that page. Installing ComfyUI is How to install ComfyUI and ComfyUI-Manager on a cloud GPU.

When it fails

Gated download. The model API reported gated: auto. Accept the license on the model page the README links, then download again. The code MIT license and the weights community license are different files.

CUDA out of memory. The tools README does not print a CUDA out-of-memory procedure. If the process raises that error, use how to fix CUDA out of memory on a cloud GPU.

“CUDA error: no kernel image is available for execution on the device” on an RTX 5090. pyproject.toml takes torch from the cu126 index on Linux x86_64. The RTX 5090 guide says that card is compute capability 12.0 (sm_120) and needs a PyTorch build for CUDA 12.8 or newer. This page did not install a different wheel and did not re-test this repository on an RTX 5090. There is still no upstream GPU GB figure to compare with the catalog’s 32 for RTX 5090 and RTX 5090D.

Save the audio, then destroy

The model card’s tools example writes output.wav in the working directory. Its diffusers example writes hammer.wav. run_gradio.py does not name an output directory. Copy the files the process writes before you destroy the instance. Save your work before you destroy a cloud GPU is the checklist. The terms say data on local instance storage is permanently deleted when the instance is destroyed.

The pricing page FAQ states the billing rule in these words:

Per-minute options have a 1-minute minimum; unused prepaid minutes are refunded when you destroy the instance. Some configurations use whole-hour billing with a 1-hour minimum; unused time within a charged hour is not refunded on voluntary stop. The billing period, minimum charge and initial prepayment are shown before confirmation. No subscription.

Prices are not copied here. The license’s US $1,000,000 and USD $1,000,000 figures above are the license text, not a price on this page.

Frequently asked questions

How much VRAM does the Stable Audio Open README or model card publish?

The stable-audio-tools README and the Stable Audio Open model card publish no GPU GB figure (read 2026-11-12). The model card says the model generates variable-length (up to 47s) stereo audio at 44.1kHz. There is no upstream GPU GB cutoff to match against vram_gb in https://cloudgpu.app/gpu-prices.json (updated_at 2026-11-12T22:09:18.532Z). This page does not mark a row as enough. Which listing you can start is on https://cloudgpu.app/deploy.

What license covers the weights, and what is the revenue threshold?

The tools repository LICENSE is the MIT License, copyright line Copyright © 2023 Stability AI. Its grant says free of charge. That sentence is the code license. It is not a CloudGPU price and not a storage allowance. The model API on 2026-11-12 reported gated auto, license_name stable-audio-community, and license_link LICENSE. LICENSE.md from that repository is titled STABILITY AI COMMUNITY LICENSE AGREEMENT, last updated July 5, 2024. Its introduction says it allows research, non-commercial, and limited commercial uses free of charge, and that it preserves free access for annual revenue of less than US $1,000,000. Those free sentences and that dollar figure are the license text. They are not a CloudGPU price and not a storage allowance. Section III says a commercial use must be registered at the URL in that section, and that the license terminates above USD $1,000,000 in annual revenue. Read the file.

How do I open the Gradio UI through SSH?

The README’s Gradio example passes --pretrained-name with a Hub repository name. Copy that argument from the README. This page does not paste the Hub id. run_gradio.py calls interface.launch with share and auth and does not pass server_name or server_port. Gradio 6.3.0, pinned by the ui extra, documents that a None server name uses 127.0.0.1 and a None server port searches for an available port starting at 7860. This page does not pass --share and does not claim the search stays on 7860. The SSH tunnel forwards 127.0.0.1 on the port the process prints.

Do the audio files stay on the instance after I destroy it?

No. The model card’s tools example saves output.wav. Its diffusers example saves hammer.wav. run_gradio.py does not name an output directory. Copy the files the process writes before you destroy the instance. The terms say data on local instance storage is permanently deleted when the instance is destroyed.

Sources (read 2026-11-12, Asia/Shanghai)

Cards you can start are on the Deploy page. Prices are on https://cloudgpu.app/pricing and are not repeated here.

Last updated: 2026-11-12 (Asia/Shanghai).