🎛️ Stable Audio 3 Local Setup: Run the Open Weights on Your GPU
Stable Audio 3's open-weight models run on consumer hardware. The Hugging Face repos, the stable-audio-3 library, license limits, and what a 4080 can expect.
Priya Nair · AI Audio & Music Producer
· 6 min read
⚡ TL;DR — quick answers
- Can you run Stable Audio 3 locally?
- Yes. Three of the four Stable Audio 3.0 models are open weights on Hugging Face: Small SFX, Small, and Medium. Small handles full music composition on-device with tracks up to two minutes, Medium generates songs past six minutes, and only Large stays locked behind the Stability AI API and enterprise self-hosting.
- What GPU do you need for Stable Audio 3?
- Stability hasn't published per-model VRAM guidance. The sizes imply a lot, though: Medium is 2B parameters (roughly 4 GB of weights at 16-bit) and both Small models are 0.6B, so a 12-16 GB consumer card has headroom. Stability pitches the Small models at phones and consumer laptops, and quotes generation in a few seconds on a MacBook Pro M4.
- Is Stable Audio 3 free for commercial use?
- Under the Stability AI Community License you own your outputs and can commercialize them as long as your organization stays under $1M in annual revenue; above that you need an Enterprise license. One extra wrinkle: the models use a T5Gemma text encoder, so Gemma Terms of Use ride along with the weights.
Three of the four Stable Audio 3.0 models went up on Hugging Face as downloadable weights on May 20, 2026, and the biggest open one is 2B parameters. That's a size a gaming GPU holds without heroics. For two years the answer to "can I run a serious music model at home" was no, and I'd gotten comfortable saying so.
Not this time.
The part I keep turning over as a musician, though, is the training set. The model cards list 1,278,902 audio recordings: 806,284 licensed from AudioSparx, 472,618 from Freesound under Creative Commons terms. Suno and Udio are still in court with major labels over exactly this question. Stability shipped the receipts alongside the weights, a sample library with its paperwork stapled to the box.
That detail decides who can safely build on it.
By the numbers
- Announced May 20, 2026; "3.0 Small SFX, 3.0 Small, and 3.0 Medium are all open-weights," per Stability's own post (stability.ai)
- Parameter counts on the Hugging Face repos: Medium at 2B, Small and Small SFX at 0.6B each (huggingface.co)
- Track length: Small composes full music up to 2 minutes; Medium goes past six minutes (Stability cites a 6:20 generation); Large also exceeds six minutes but stays API-only
- Training data per the model cards: 1,278,902 recordings, split 806,284 AudioSparx / 472,618 Freesound (266,324 CC-0, 194,840 CC-BY, 11,454 CC-Sampling+)
- Speed, per Stability's research note: music and sounds in under about 2 seconds on an H200 and "a few seconds" on a MacBook Pro M4, with quickstart examples running just 8 inference steps
- License: Stability AI Community License, free until your org passes $1M annual revenue, then Enterprise
- Per-model VRAM guidance: not published as of this writing
What's actually in the family
| Model | Params (per HF) | Weights | What it's for |
|---|---|---|---|
| 3.0 Small SFX | 0.6B | Open | Sound effects; Stability aims it at phones and consumer laptops |
| 3.0 Small | 0.6B | Open | Full music composition on-device, tracks up to 2 minutes |
| 3.0 Medium | 2B | Open | Full songs with structure past six minutes |
| 3.0 Large | Not published | Closed | Stability AI API and enterprise self-hosting only |
The one spec I couldn't pin down is a maximum duration for Small SFX; the card talks about variable-length generation without committing to a ceiling. For sound design that rarely matters. A whoosh is four seconds long.
All three open models generate stereo, and all three support inpainting: regenerate one region of a track, or continue a short recording, without re-rolling the rest. That's the feature I'd actually test first, because section-patching is where AI music tools either earn a place in a real workflow or get uninstalled by Friday.
Getting the weights running
The weights are gated, so step one is accepting the license terms on each Hugging Face repo while logged in. After that, Stability points at two stacks. The new stable-audio-3 library is the inference-and-fine-tuning path, a from_pretrained("medium") call and you're generating. The older stable-audio-tools research library also works, and its example configs run 8 inference steps, which is why the speed claims aren't fantasy: adversarial post-training cut the step count down hard.
There's no Diffusers pipeline in the official instructions, so don't plan around one. ComfyUI, on the other hand, was named a day-one partner in the announcement, and node graphs are honestly a good fit for audio inpainting. If you already have a ComfyUI install for image work, this bolts on.
One licensing footnote most coverage skipped: the text encoder is T5Gemma, so the Gemma Terms of Use apply alongside Stability's Community License. Two license texts, one model folder. Read both once.
What a consumer GPU can expect
Stability published no VRAM table, so here's what the arithmetic implies instead. Medium's 2B parameters at 16-bit precision come to roughly 4 GB of weights before activations and the text encoder; the 0.6B Smalls sit near 1.2 GB. On our RTX 4080, a 16 GB card, that leaves room to spare for either model, and Stability's own framing agrees: the Smalls are pitched at phones and consumer laptops, with full music composition on-device the stated point of Small.
What I won't do is invent a tokens-per-second equivalent for audio. Stability's published numbers are an H200 generating in under two seconds and a MacBook Pro M4 taking a few seconds; a 4080 lands somewhere in that range on the small models, and we'll publish real timings in a follow-up once we've run the same prompts across all three checkpoints. Vendor demo speeds have burned me before. (Every "real-time" TTS claim I've tested had an asterisk the size of a verse.)
What I'll be listening for isn't speed anyway. Small diffusion models fail audibly in specific places: cymbal tails that dissolve into static, sustained pads that develop a shimmer, a bass line that loses pitch center in the last chorus. If the 0.6B models keep those artifacts out of a two-minute track, the size stops being a compromise.
The license is doing real work here
The Community License terms are unusually friendly for weights this capable: you own your outputs and can distribute and commercialize them freely until your organization crosses $1M in annual revenue, at which point Stability wants an Enterprise deal. For an indie composer, a YouTube channel, or a game studio below that line, the math is generation at the cost of electricity.
Compare that with the cloud subscription tools in our best AI music generators rundown, where the monthly bill scales with output, or the capped trials in the free AI music generators list. Open weights on your own card is the only version of "free" without a meter attached.
And the provenance angle is bigger than a compliance checkbox. When a client asks where the music came from, "a model trained on 806,284 licensed AudioSparx recordings plus Creative Commons Freesound audio" is an answer with numbers in it. The Suno vs. ElevenLabs Music comparison covers how the litigation risk shakes out between the big cloud platforms; Stable Audio 3 sidesteps the fight by publishing its sources. Whether the licensed-only diet costs it some raw musicality is exactly what a listening test should answer.
How we're testing
Everything above traces to Stability's May 20 announcement, its research note, and the Hugging Face model cards; the figures we couldn't source from those pages are marked not published, and the Stable Audio tool page tracks the hosted version separately. The follow-up plan: all three open checkpoints on the 4080, identical prompts, with the inpainting seams and the last ten seconds of every long generation under headphones.
Open weights mean the model's flaws become public knowledge instead of a changelog entry. That trade favors anyone who listens closely.
Frequently asked questions
▸Can you run Stable Audio 3 locally?
Yes. Three of the four Stable Audio 3.0 models are open weights on Hugging Face: Small SFX, Small, and Medium. Small handles full music composition on-device with tracks up to two minutes, Medium generates songs past six minutes, and only Large stays locked behind the Stability AI API and enterprise self-hosting.
▸What GPU do you need for Stable Audio 3?
Stability hasn't published per-model VRAM guidance. The sizes imply a lot, though: Medium is 2B parameters (roughly 4 GB of weights at 16-bit) and both Small models are 0.6B, so a 12-16 GB consumer card has headroom. Stability pitches the Small models at phones and consumer laptops, and quotes generation in a few seconds on a MacBook Pro M4.
▸Is Stable Audio 3 free for commercial use?
Under the Stability AI Community License you own your outputs and can commercialize them as long as your organization stays under $1M in annual revenue; above that you need an Enterprise license. One extra wrinkle: the models use a T5Gemma text encoder, so Gemma Terms of Use ride along with the weights.
▸How is Stable Audio 3 different from Suno and Udio?
Training data. Stability lists exactly what the models learned from: 806,284 recordings licensed from AudioSparx plus 472,618 Creative Commons recordings from Freesound. Suno and Udio face active copyright litigation from major labels over their training sets, which matters if you plan to monetize what you generate.
The 5 best AI video finds, every week
New models, tested prompts, and what actually worked in our production — one short email a week. No spam, unsubscribe anytime.
Written by Priya Nair
AI Audio & Music Producer
Musician first, AI tester second. Judges every voice and song model on whether it would survive a real listener's headphones, not a spec sheet — and keeps a close eye on what's actually safe to monetize.
Explore these topics
Every guide, comparison and prompt library we have on each.
Keep learning
GPUs & Hardware · Local Video
GuidesRun Stable Diffusion Locally: Setup That Works in 2026
SD1.5 vs SDXL VRAM math, the ComfyUI desktop install, your first workflow, where checkpoints come from, and the errors that eat everyone's first afternoon.
