Self-hosted vs. API generative audio: choosing a deployment model
September 9, 2026
Self-hosted generative audio runs model weights on your own hardware, trading a fixed compute cost and full data control for GPU management. An audio generation API removes that operational load and bills per generation. Stable Audio 3.0 offers both: three open-weight models, a managed API, and an enterprise self-hosted license.
Key takeaways
Stable Audio 3.0 Small-Music and Small-SFX run text-to-audio on CPU with no GPU at all.
Stable Audio 3.0 Medium peaks at 6.52 GB of VRAM, so it fits on a single mid-range consumer GPU.
The largest model in the family, Stable Audio 3.0 Large, sits outside the open-weight release. Reaching it means the API or an enterprise self-hosted license.
Legal indemnification is offered as part of the Enterprise license, not to Community License self-hosting.
Three deployment routes for generative audio, compared
Generative audio deployment splits three ways, and the middle route is the one most comparisons skip. Open-weight self-hosting puts the model on your hardware under a free license. An enterprise self-hosted license puts a larger model on your hardware under a paid contract with legal cover. A managed API puts nothing on your hardware at all.
Stability AI publishes all three for the Stable Audio 3.0 model family. Three of the four models, Small, Small-SFX and Medium, are downloadable from Hugging Face under the Stability AI Community License and free to run commercially below $1M in annual revenue. Stable Audio 3.0 Large runs through the Stability AI developer platform, and enterprise customers can deploy it inside their own infrastructure under a paid license.
Most teams evaluating generative audio compare hosted products against each other and never reach this question. The route you pick determines your cost curve, your data boundary and your legal exposure, and those three things are harder to change later than the model itself.
| Route | How you get it | Models available | Cost model | Indemnification |
|---|---|---|---|---|
| Open-weight self-hosting | Download from Hugging Face, Community License | Small, Small-SFX, Medium | Your compute, no license fee below $1M revenue | No |
| Enterprise self-hosted license | Paid contract with Stability AI | Full family, including Large | Negotiated license plus your compute | Under enterprise terms |
| Stability AI API | API key, pay per generation | Large | Per generation | Under enterprise terms |
Self-hosted generative audio: running the weights yourself
Self-hosting converts a per-generation bill into a fixed infrastructure cost and an operational responsibility. You download the weights, stand up an inference environment, and absorb everything the vendor was previously handling: capacity, uptime, dependency management, and the upgrade path when a new model ships.
What you get back is control. Prompts and generated audio never leave your network. Generation volume stops mapping to spend. You have maximum control over fine-tuning, and Stable Audio 3.0 ships LoRA training documentation alongside the open weights, so adapting a model to a house style or a client's back catalog is a supported workflow rather than a hack.
What hardware Stable Audio 3.0 needs to run locally
Stable Audio 3.0 Medium peaks at 6.52 GB of VRAM generating a 380-second track, which puts the best open-weight model in the family inside a single mid-range consumer GPU. Stable Audio 3.0 Medium is a 1.4B-parameter model requiring CUDA and Flash Attention 2. Without Flash Attention 2 correctly installed, generation degrades to unusable output rather than failing cleanly, which is worth knowing before you debug the model instead of the environment.
Small and Small-SFX need no GPU whatsoever. Both are 433M parameters, both generate up to 120 seconds, and both are built for CPU inference on consumer laptops and mobile hardware.
Published runtimes cover CPU via TFLite, Apple Silicon via MLX, and NVIDIA via TensorRT, so the hardware floor is a laptop rather than a rented instance.
| Model | Parameters | Hardware | Max length | Peak VRAM | Generation time (H200) |
|---|---|---|---|---|---|
| Small-Music | 433M | CPU | 120s | 2.40 GB | 0.45s |
| Small-SFX | 433M | CPU | 120s | 2.40 GB | 0.45s |
| Medium | 1.4B | GPU (CUDA) | 380s | 6.52 GB | 1.31s |
| Large | 2.7B | API or enterprise self-hosted | 380s | 8.01 GB – 9.04 GB | 1.80s on H200 |
Figures from the Stable Audio 3.0 repository. VRAM is peak allocated on H200 with unchunked decode. Chunked decoding lowers it: Medium at 120 seconds drops from 6.49 GB to roughly 5.14 GB. The architecture behind those numbers is documented in the Stable Audio 3.0 technical report.
Who owns the output when you self-host an audio model
As between you and Stability, you own your outputs, regardless of whether you self-host the model or call our APIs, and whether you're using the base models or a fine-tune.
One asymmetry matters for procurement. The Community License grants rights; it doesn't grant indemnity. Legal indemnification for generated audio is an Enterprise license feature, and self-hosting under the free license leaves you carrying that risk yourself.
API generative audio: what a managed endpoint gives you
An audio generation API buys three things self-hosting can't: the largest model in the family, zero hardware procurement, and someone else's problem when demand spikes. Stable Audio 3.0 Large is a 2.7B-parameter model best for highest-quality generation and is the slowest of the models we offer. It's reachable through the Stable Audio API reference without a contract negotiation.
Time to first generation is the underrated part. An API key works this afternoon. A self-hosted Medium deployment means CUDA versions, a Flash Attention 2 wheel matching your Python and PyTorch build, and a GPU you may not have yet.
How self-hosting and API access differ on data control and indemnification
Self-hosting keeps prompts and generated audio inside your own infrastructure, and the Enterprise license is what adds legal cover on top. Those two protections come from different places, which is why the strongest compliance position combines them rather than choosing between them.
For teams working under client confidentiality, unreleased-project secrecy or a data residency requirement, the argument is straightforward. Locally run weights mean no prompt text and no audio crosses a network boundary. Nothing is logged by a third party, because no third party is involved.
Indemnification runs the other way. Stability AI trained the Stable Audio 3.0 family on fully licensed data, and offers legal indemnification under the Enterprise license. Community License self-hosting inherits the benefits of licensed training data without the indemnity. An organization that needs both privacy and contractual cover is looking at an enterprise self-hosted deployment, which is precisely the route the two-way framing of this question tends to hide.
How to choose between self-hosting and an audio generation API
Volume, hardware access and the quality ceiling you need decide this, roughly in that order. Everything else is secondary.
| Situation | Route | Reason |
|---|---|---|
| Prototyping with no GPU available | Small on CPU, or the API | Neither requires hardware procurement to start |
| Low or irregular monthly volume | API | Fixed hardware cost spreads across too few generations |
| Steady high volume, GPU available, instrumental tracks up to 6:20 | Medium, self-hosted | Per-generation cost approaches electricity |
| Audio or prompts cannot leave your infrastructure | Self-hosted | No network boundary is crossed |
| Highest musicality in the family, plus indemnification | Large via API or enterprise license | Large is outside the open-weight release |
A pattern worth noticing across those rows: the routes aren't mutually exclusive. Prototyping on Small locally, shipping on the API, and moving to self-hosted Medium once volume justifies the hardware is a sequence within one model family, on one license, with one set of output rights. Switching between them costs engineering time. It doesn't cost a vendor migration, a new prompt library, or a fresh negotiation over who owns the audio.
For a model-by-model view of what each variant produces, see our comparison of Stable Audio against other generative audio platforms.
Frequently asked questions about self-hosted and API audio generation
Can you self-host Stable Audio 3.0? Yes. Stability AI released three of the four Stable Audio 3.0 models as open weights: Small, Small-SFX and Medium. Download them from the Stable Audio 3.0 collection on Hugging Face and run them on your own hardware under the Community License. Stable Audio 3.0 Large falls outside that release.
What GPU do you need to run Stable Audio 3.0? Stable Audio 3.0 Medium peaks at 6.52 GB of VRAM generating a 380-second track on an H200, which fits on a single mid-range consumer card. Medium also requires Flash Attention 2 and a CUDA environment. Small and Small-SFX need no GPU.
Can Stable Audio 3.0 run without a GPU? Yes. Small and Small-SFX are built for CPU inference and generate up to 120 seconds of audio.
Is self-hosting Stable Audio 3.0 free? The weights and the license are free below $1M in annual revenue. Compute is not. Self-hosting replaces per-generation billing with hardware cost plus the engineering time to keep an inference environment running. Above $1M in revenue, an Enterprise license applies.
Do you own the music generated by a self-hosted audio model? Under the Stability AI Community License, you own outputs from the Core Models and from fine-tunes built on them. Ownership terms don't differ between the self-hosted and API routes. All use of Stability AI models is subject to the Stability AI Acceptable Use Policy.
Is Stable Audio 3.0 Large available for self-hosting? Yes, though not through the open-weight release. Large runs on the Stability AI API, and self-hosted deployment is available to enterprise customers under a paid license. Teams wanting the family's highest musicality on their own infrastructure need that contract.
Does Stability AI offer indemnification for self-hosted deployments? Legal indemnification attaches to the Enterprise license rather than to Community License use. Self-hosting under the free license gives you the weights and output ownership without an indemnity. Organizations needing contractual cover for generated audio should evaluate the Enterprise route.
Last updated September 9, 2026
