What is Stable Audio 3.0 Medium?
Stable Audio 3.0 is a family of four text-to-audio models from Stability AI, announced on 20 May 2026. Three have open weights: Medium, for music up to 6:20 with stronger structure; Small (published as Small-Music), for music up to two minutes on device; and Small SFX, for sound effects and ambience on phones and laptops. Large is available only through Stability's API or an enterprise deployment.
Stability says all four are trained on licensed data; the model card lists 806,284 recordings from the AudioSparx library and 472,618 Creative Commons recordings from Freesound. Medium has about 2.3 billion parameters including its T5Gemma text encoder, and each small model about 568 million, per the Hugging Face file listings. Length is set to the second, and inpainting edits one or several segments or continues a recording.
The weights are under the Stability AI Community License: free, commercial use included, for organisations under $1 million in annual revenue.
Is Stable Audio 3.0 Medium worth it?
Stable Audio 3.0 is the open model family for instrumental music and sound effects where provenance matters. Stability trained it on 1,278,902 licensed recordings from AudioSparx and Freesound and screened out copyrighted music with an audio tagger and an outside detection company, according to the model card. No other open music model we cover documents its data as precisely.
Medium, the largest open model, writes structured instrumental tracks up to 6:20, with length set to the second. The model card reports generation in under 2 seconds on an H200 and a few seconds on a MacBook Pro M4. The two small models, about 568 million parameters each, are built for phones and laptops: Small-Music makes up to two minutes of music, up from 11 seconds for Stable Audio Open Small, and Small SFX makes sound effects and ambience. Inpainting regenerates chosen segments or extends a track, and LoRA training is documented for Small and Medium.
The catch: this is not a song model. The input is a text prompt with no lyrics field, and Stability does not claim sung vocals, so anyone who needs a singer should look at ACE-Step 1.5 or MiniMax Music 3. The Hugging Face weights are gated behind a licence agreement, the Community License stops being free at $1 million in annual revenue, and the bundled T5Gemma encoder brings the Gemma Terms of Use with it. The strongest model, Large, cannot be downloaded at all.
Our take: for game and app studios that need sound effects or background music generated on the device, with a clean answer when the legal team asks about training data, Stable Audio 3.0 is the natural open pick; for songs with words it is the wrong tool.
Where it earns its place
A mobile game studio under $1 million in revenue
Generate sound effects on the player's device instead of shipping large audio files
Bundle Small SFX in the app and generate effects on demand; the Community License covers commercial use below $1 million a year, after which an Enterprise licence is needed.
A podcast network's production team
Produce intros, music beds and transitions with a clear rights trail
Generate instrumentals with Medium at lengths set to the second and use inpainting to rework a single segment; the documented AudioSparx and Freesound training data answers clients' provenance questions.
A developer prototyping without a GPU
Test music generation in an app before committing to self-hosting
Call Stable Audio 3.0 through Stability's API at 8 credits ($0.08) per generation; the 25 free credits on a new account cover three generations.
Before you commit
- Setup
- Accept the licence on Hugging Face, install the stable-audio-3 library and call StableAudioModel.from_pretrained("medium") with a prompt and a duration; ComfyUI had support on launch day.
- Integrations
- stable-audio-3 and stable-audio-tools Python libraries, ComfyUI, the Stability AI API, and hosted inference through fal.
- Data and privacy
- Self-hosted weights keep audio local. Training data: 806,284 AudioSparx recordings and 472,618 Freesound recordings under CC-0, CC-BY or CC Sampling+. Under the Community License you own your outputs.
- Cost at scale
- Free to self-host for organisations under $1 million revenue; above that an Enterprise licence at custom pricing, which adds legal indemnification. On the API, 10,000 generations at $0.08 cost $800.
- Limits
- Medium up to 6:20, small models up to two minutes; English prompts; no lyrics or vocal control; the Gemma Terms of Use restrictions apply to the text encoder.
What’s good
- Trained on 1,278,902 licensed recordings, with copyright screening
- Medium makes up to 6:20, length set to the second
- Small models (about 568M parameters) run on phones and laptops
- Small SFX covers sound effects and ambience
- Inpainting, continuation and LoRA training
What’s not
- No lyrics input; not a vocal song model
- Free licence ends at $1 million annual revenue
- Gated download; Gemma terms also apply
- Large, the best model, is API or enterprise only
- Small models cap at two minutes
Best for
Game, app and video teams that need instrumental music or sound effects from a model with documented, licensed training data, including on device.
Stable Audio 3.0 Medium alternatives
Stable Audio 3.0 Medium: common questions
- Is Stable Audio 3.0 Medium worth it?
- Stable Audio 3.0 Medium is the open choice for instrumentals and sound effects trained on licensed audio, with small siblings that run on a phone, but it does not make songs with lyrics. We rate it 3.9 out of 5.
- How much does Stable Audio 3.0 Medium cost?
- Weights free under the Stability AI Community License for organisations under $1 million annual revenue; above that, an Enterprise licence at custom pricing. API: Stable Audio 3.0 costs 8 credits ($0.08) per generation; new accounts get 25 free credits.
- What are the downsides of Stable Audio 3.0 Medium?
- No lyrics input; not a vocal song model; free licence ends at $1 million annual revenue; gated download; Gemma terms also apply; large, the best model, is API or enterprise only; small models cap at two minutes.
- Who is Stable Audio 3.0 Medium best for?
- Game, app and video teams that need instrumental music or sound effects from a model with documented, licensed training data, including on device.
- How much does Stable Audio 3.0 Medium cost for a team or in production?
- Free to self-host for organisations under $1 million revenue; above that an Enterprise licence at custom pricing, which adds legal indemnification. On the API, 10,000 generations at $0.08 cost $800.
- Is Stable Audio 3.0 Medium safe for company data?
- Self-hosted weights keep audio local. Training data: 806,284 AudioSparx recordings and 472,618 Freesound recordings under CC-0, CC-BY or CC Sampling+. Under the Community License you own your outputs.
- What are the best Stable Audio 3.0 Medium alternatives?
- The closest alternatives we have reviewed are ACE-Step 1.5, Eleven Music v2.5 and Lyria 3.5.
Answers checked Nov 7, 2026.