Meet Stable Audio 3.0 — the model family for artistic experimentation with open-weight models

TL;DR

Stability AI announces the Stable Audio 3.0 family — four models (Small SFX, Small, Medium, Large) covering on-device sound-effects through full-song composition at up to ~6:20. Three of the four (Small SFX, Small, Medium) ship as open weights on Hugging Face under the Stability AI Community License; Large is available via the Stability API and enterprise licensing. All four are trained only on licensed data, support variable-length generation at per-second granularity, support segment/causal-continuation inpainting, and ship with first-party LoRA-training documentation. The underlying architecture (semantic-acoustic autoencoder + latent diffusion + adversarial post-training) is documented in the companion arxiv paper Stable Audio 3.

Key claims

Stable Audio 3 is released as a four-model family: Small SFX (on-device SFX), Small (full music composition on-device), Medium (higher musicality up to 6:20), Large (most advanced musicality, low-latency at high volume) [§Meet the Stable Audio 3.0 model family].
Three of the four — Small SFX, Small, Medium — are released as open weights on Hugging Face; Large is available only via the Stability AI API and self-hosting under an enterprise license [§Key Takeaways, §Get started].
All four models are trained on fully licensed data only [§Key Takeaways, §Open for experimentation].
The Stability AI Community License gives users ownership of their outputs and the right to distribute and commercialize them, until the user organization passes a $1M revenue threshold — above which an Enterprise license (with legal indemnification) applies [§Open for experimentation].
Stable Audio 3 introduces a new method for variable-length audio generation at per-second granularity, replacing the previous practice of always generating a fixed-length clip [§Architectural advancements].
Small generates up to 2 minutes, vs 11 s for Stable Audio Open Small and 47 s for Stable Audio Open; Medium and Large generate more than six minutes [§Architectural advancements].
Small is positioned as the only model capable of full music composition on-device — i.e. complete tracks, not just short samples, offline on consumer hardware [§Architectural advancements].
LoRA training is documented as a first-party capability alongside the weights for Small and Medium — the blog frames this as the first time Stability is publishing LoRA-training documentation for an audio model [§Architectural advancements].
Inpainting in Stable Audio 3 covers three modes: single-segment editing, multi-segment editing, and causal continuation (extending audio beyond its original endpoint) [§Architectural advancements].
Stable Audio 3 is the architecture base for Stability’s next-generation licensed audio products for professionals; partnerships with Universal Music Group and Warner Music Group are explicit framing for the licensed-platform value proposition [§Setting the stage].
Day-0 ComfyUI integration ships alongside the API and Hugging Face release [§Get started, ComfyUI blog].

Method

The blog itself is a product/license announcement, not a technical paper — the architectural details (semantic-acoustic autoencoder, latent diffusion, adversarial post-training) live in the companion arxiv report Stable Audio 3. What this blog adds is the productization layer:

Variant grid: SFX-only Small vs music-capable Small (both on-device-targeted), Medium and Large as cloud/larger-compute tiers. The SFX-only Small variant is the new shape vs prior Stable Audio Open releases — Stability is segmenting the family by application (SFX vs music composition) at the smallest size before scaling.
License tiering: Community License (open weights, commercial use under revenue threshold), Enterprise License (legal indemnification, fine-tuning support, available above the threshold), and API-only access for Large. The license tier maps to model size.
First-party LoRA pipeline: “documentation for LoRa training, alongside the weights for 3.0 Small and 3.0 Medium” — explicitly framed as an emerging method for audio (well-established for images via Stable Diffusion) that the team is now formalizing for music.
Distribution: Hugging Face collection for open-weights models; Stability API for Large; ComfyUI as a day-0 third-party platform.

The launch positions this release explicitly alongside the company’s licensing partnerships with UMG and WMG, which is the strategic framing — “responsibly trained generative AI” as a competitive lever vs unlicensed competitors — rather than purely a technical message.

Results

No benchmarks in the blog itself. Operational claims:

Small variants run on-device on consumer-grade laptops and mobile phones.
Variable-length generation at per-second granularity, with Small reaching 2 minutes (up from 11 s / 47 s on prior Stable Audio Open releases) and Medium/Large reaching beyond 6:20.
Generation latency claims (<2 s on H200, “a few seconds on MacBook Pro M4”) are in the companion arxiv paper, not in the blog body.

Why it’s interesting

The release shape is a useful new datapoint for Open foundation-model releases: a four-model family with deliberate application segmentation at the smallest tier (Small SFX vs Small for music) and license-tiered open access (three open + one API-only) is a structure the wiki hasn’t filed before. ACE-Step 1.5 (ACE-Step 1.5: Pushing the Boundaries of Open-Source Music Generation) ships a similar variant grid by size and turbo/SFT axis, but every checkpoint is open MIT; LongCat-AudioDiT (LongCat-AudioDiT: High-Fidelity Diffusion Text-to-Speech in the Waveform Latent Space) ships 1B + 3.5B both MIT. Stable Audio 3’s design is closer to the Mistral / DeepSeek pattern: open up to a size cap, hold the largest model as a hosted product, and use the open variants to drive ecosystem and the closed variant to capture revenue.

The first-party LoRA-training documentation for an audio diffusion model is also worth flagging for Parameter-Efficient Finetuning: LoRA on audio has existed in the community for a while, but treating it as a shipping artifact (docs + weights co-released so finetuning is the documented happy path) is the same productization step that image-generation LoRAs went through three years ago. If the audio community follows the image pattern, the next 6–12 months will produce a Stable-Audio-3 LoRA marketplace analogous to the SD ecosystem — a real Luma-relevant signal if Luma ships audio at any point.