Stability AI has released Stable Audio 3.0, a family of four audio generation models that can produce coherent music tracks running past six minutes — more than double the roughly 190-second ceiling of Stable Audio 2.5, the company’s own prior internal baseline, according to its research paper. Three of the four models ship with open, downloadable weights. The fourth does not.
That distinction matters more than the headline number. Stable Audio 3.0 isn’t one model with one set of capabilities — it’s four models with different hardware requirements, different maximum lengths, and different rules about who can download what. Understanding which model does what is the difference between a useful evaluation and a misleading one.
Table of Contents
What Stability AI Announced
Stability AI published the Stable Audio 3.0 announcement on May 20, 2026, describing it as a model family “built for artistic experimentation,” trained entirely on licensed and Creative Commons audio. The pitch: longer generations, an updated architecture, support for fine-tuning via LoRA, and a licensing structure the company says lets users own and commercialize their output.
The release landed the same day as Day-0 support in ComfyUI, the popular node-based interface for generative media, and followed a research paper published two days earlier detailing the underlying architecture.
The Four Models, and What Each One Actually Does
The family splits into two tiers. Small SFX and Small (sometimes labeled Small-Music) are lightweight, CPU-only models capable of about two minutes of audio. Medium and Large are the models capable of the six-minute-plus generations that anchor the announcement — and only Medium’s weights are publicly available. Large is reachable exclusively through Stability’s API or an enterprise self-hosting agreement; the company has not stated when, or whether, that will change.
| Model | Weights | Hardware | Max length | Primary use |
|---|---|---|---|---|
| Small SFX | Open (Hugging Face) | CPU only | ~2:00 | Sound effects |
| Small (Music) | Open (Hugging Face) | CPU only | ~2:00 | Short music, on-device composition |
| Medium | Open (Hugging Face) | GPU (CUDA) | ~6:20 | Full-length music, higher musicality |
| Large | API / enterprise self-host only | API-side | ~6:20 | Highest quality, high-volume use |
This means the “six-minute generation” that headlines the release is a Medium and Large capability, not a family-wide one. And the “on-device” composition Stability highlights — full songs generated locally, without a cloud connection — describes the Small models specifically. Medium needs an actual GPU.
The New Architecture Behind It
Stable Audio 3.0 runs on a newly designed autoencoder, which Stability’s research team calls SAME — a Semantically-Aligned Music Autoencoder, detailed in a companion paper alongside the main technical report. The autoencoder compresses audio into a compact latent space that the diffusion model then generates from, working in stereo at 44.1kHz with 256-dimensional latents.
The practical effect, according to Stability, is that the models can generate several minutes of coherent audio without the runaway compute cost that would come from generating that much raw audio directly. It’s the same broad idea — compress, generate in the compressed space, then decode — that underlies most modern diffusion-based image and audio systems, applied here with an autoencoder built specifically to preserve musical structure over longer spans.
LoRA Fine-Tuning and Customization
For creators who want to adapt these models to their own sound, Stability has published an actual LoRA training workflow alongside the weights — not just a mention of the feature, but documentation, code, and a training path built for Apple Silicon Macs using the MLX framework, in addition to the standard GPU-based route. LoRAs can be loaded at inference time, stacked, and adjusted in strength without retraining the base model.
This applies to Small and Medium, the two models actually distributed through Stability’s open-source repository. Large isn’t supported by that codebase at all, so there’s no public LoRA path for it.
Training Data and Licensing Claims
Stability has been more specific than usual about where the training data came from. According to the company’s own model documentation, the training set includes just over 1.27 million audio recordings: roughly 806,000 licensed from the stock-audio library AudioSparx, and about 472,000 from Freesound, a Creative Commons audio repository. The Freesound portion was filtered using an automated music-detection tool and reviewed by an outside content-verification firm to remove anything that might contain unlicensed copyrighted material.
That’s a meaningfully more transparent accounting than most AI audio releases offer, and it’s a direct response to the copyright disputes that have dogged competitors. It’s worth noting, though, that this is Stability’s own disclosure — it hasn’t been independently audited by a third party as far as public records show.
What “Open Weights” Actually Means Here
“Open weights” is doing some real work in Stability’s marketing, and it’s worth unpacking. Downloading Small SFX, Small, or Medium from Hugging Face requires agreeing to Stability’s Community License and providing contact information — it’s not an anonymous, no-strings download. Under that license, users own their generated output and can commercialize it freely, but only as long as their organization’s annual revenue stays under $1 million. Above that threshold, a separate, paid Enterprise License is required, which adds legal indemnification and optional support for custom fine-tuning.
There’s another layer specific to the Medium model: it uses a Google-built text-understanding component called T5Gemma for interpreting prompts, which is redistributed under Google’s own Gemma Terms of Use — a separate set of restrictions layered on top of Stability’s license. None of this is disclosed in Stability’s original announcement post; it only appears in the more detailed technical documentation on Hugging Face.
None of this makes the release less open than comparable efforts in the space. But “free to use commercially” and “free to use commercially if your organization earns under $1 million a year, under two overlapping licenses” are different claims, and the latter is the accurate one.
Hardware Requirements and Performance
The Small models are built to run on ordinary hardware — no GPU required, low resource use, suitable for laptops or even mobile-class devices. Medium is a different proposition. Stability’s own repository specifies that Medium requires an NVIDIA GPU with CUDA support and a library called Flash Attention 2 correctly installed; without it, the documentation notes, output can come out as unusable static. Reported memory use for Medium ranges from roughly 5 to 6.5 gigabytes of GPU memory depending on the length of the generated clip.
There’s a small inconsistency worth flagging in Stability’s own materials. One piece of documentation states that the model family generates audio “in less than a few seconds on a MacBook Pro M4” — language that reads as applying broadly. But the company’s own published performance benchmarks only include Mac-based timing figures for the Small model. Medium’s benchmarks are GPU-only. So while optimized Mac and CPU inference paths exist for parts of the family, it’s not clear from the public record that Medium has been benchmarked running well on a Mac specifically, despite that one broader claim.
Availability — Where to Get It
Small SFX, Small, and Medium can be downloaded now from Stability AI’s Hugging Face page, subject to the license terms described above. Large is accessible only through Stability’s developer API, billed on a credit system, or through a custom enterprise self-hosting arrangement. Exact per-generation API pricing for Large has not been confirmed through Stability’s own official pricing documentation; third-party estimates exist but disagree with each other, so no specific figure can be reported as verified.
ComfyUI added support for the Small and Medium models on launch day, letting users run them inside existing node-based generation workflows rather than through a separate app.
How It Compares
Context matters here more than a head-to-head score. Stability’s own research paper compares Stable Audio 3.0 against its prior internal model, Stable Audio 2.5, reporting higher scores on a company-run musicality listening test. That’s a meaningful internal improvement, but it’s a company-run comparison, not independent verification.
Consumer tools such as Suno and Udio occupy a different part of the market — built around hosted, ready-to-use generation rather than self-hosted, fine-tunable open models. This report’s evidence is scoped to Stability AI’s own release, so no verified claim is made here about how Stable Audio 3.0 compares to those tools on output quality, feature set, or legal status. What can be said with confidence is that no independent, third-party benchmark currently exists comparing Stable Audio 3.0 to any competing music-generation model.
What’s Still Unverified
A few specifics remain genuinely unresolved. Stability’s own sources don’t agree on the models’ exact parameter counts — the official GitHub repository, the Hugging Face model listings, and early press coverage each report different figures for the same models, sometimes by a wide margin. Precise API pricing for the Large model is similarly unconfirmed by an official source. And as noted above, no independent organization has published a comparative quality benchmark for Stable Audio 3.0 against rival tools. These aren’t small oversights in the coverage — they’re gaps in what Stability itself has made public.
Later Developments
In August 2026, Stability AI extended the Stable Audio 3.0 ecosystem with a plugin that runs generation directly inside digital audio workstations — named support so far includes Ableton Live and Logic Pro — alongside a rebuilt version of its consumer web app, StableAudio.com, focused on iterating on a track rather than generating once and moving on. Both launched as public betas, separate from the original May release, with Stability describing them as early and still experimental.
Conclusion
Stable Audio 3.0 is a genuine technical step forward for open-weight audio generation, and the transparency around training data is a real point of differentiation in a field shadowed by copyright litigation. But the release rewards reading the fine print. The six-minute generations, the on-device composition, and the “open” weights each apply to specific models under specific conditions — not uniformly across the family. Whether Stable Audio 3.0 becomes a serious tool for musicians and developers will likely depend less on the launch-day claims and more on questions the public record doesn’t yet answer: how it actually sounds compared to competitors, what the Large model will eventually cost at scale, and whether Stability opens further weights over time.
FAQ
Is Stable Audio 3.0 free to use commercially? Yes, for the open-weight models (Small SFX, Small, and Medium), as long as your organization’s annual revenue is under $1 million. Above that threshold, a separate paid Enterprise License is required.
What’s the difference between Stable Audio 3.0 Small, Medium, and Large? Small models run on CPU only and generate up to about two minutes of audio, aimed at lightweight or on-device use. Medium requires a GPU and can generate over six minutes with stronger musical structure. Large offers the highest quality but is only accessible through Stability’s API or enterprise self-hosting.
Can I download the Stable Audio 3.0 Large model? No. Large’s weights are not publicly available, and Stability has not announced a timeline for changing that.
Does Stable Audio 3.0 support LoRA fine-tuning? Yes, for the Small and Medium models, with published training documentation and a dedicated workflow for Apple Silicon Macs. Large is not supported by Stability’s open-source training code.
What hardware do I need to run Stable Audio 3.0 Medium? An NVIDIA GPU with CUDA support and the Flash Attention 2 library installed. Stability’s documentation reports GPU memory use of roughly 5 to 6.5 gigabytes depending on the length of the output.
What data was Stable Audio 3.0 trained on? According to Stability, about 1.27 million audio recordings: roughly 806,000 licensed from AudioSparx and about 472,000 Creative Commons recordings from Freesound, screened to remove unlicensed music content.
Is Stable Audio 3.0 available in ComfyUI? Yes. ComfyUI added support for the Small and Medium models on the same day Stable Audio 3.0 launched.