Stable Audio 3.0 is Stability AI’s May 2026 family for generating and editing music and sound. Its most important change is distribution: three checkpoints have downloadable weights, including models small enough to run on a CPU and a larger model intended for a CUDA GPU.
“Open weights” is the accurate term. The inference code is published in a public repository, but the model checkpoints are governed by the Stability AI Community License rather than an unrestricted open-source software license. That distinction matters for commercial adoption.
Four models, three weight releases
The family covers different audio jobs and deployment sizes:
| Model | Parameters | Maximum length | Intended hardware | Availability |
|---|---|---|---|---|
| Small Music | 433M | 120 seconds | CPU or optimized portable runtime | Downloadable weights |
| Small SFX | 433M | 120 seconds | CPU or optimized portable runtime | Downloadable weights |
| Medium | 1.4B | 380 seconds | CUDA GPU | Downloadable weights |
| Large | 2.7B | 380 seconds | Hosted or enterprise infrastructure | API; not in community repo |
Stability’s launch article describes Large as available through the Stability API and enterprise self-hosting. The official repository labels it API-only because the community code does not download that checkpoint. Teams considering enterprise deployment should confirm the exact delivery and license with Stability AI.
Small Music and Small SFX share a compact architecture but have different post-training. Use the music checkpoint for composition and the SFX checkpoint for discrete sounds rather than expecting one small model to excel at both. Medium targets higher musicality, structure, phrasing, and longer tracks.
The SAME autoencoder
Stable Audio 3 generates in a compressed audio space. Its new Semantic-Acoustic Music Encoder, or SAME, converts 44.1 kHz stereo audio into 256-dimensional latent representations. A diffusion transformer denoises those latents under text and duration conditioning, and the decoder reconstructs a waveform.
The name reflects two goals. The latent should retain acoustic detail—timbre, texture, stereo information—and also preserve higher-level musical organization. A good reconstruction codec is not enough if the latent space makes long structure difficult to learn.
Stability also uses adversarial post-training to reduce inference steps while improving fidelity and prompt adherence. Published performance figures are measured on specific hardware and optimized runtimes; they should not be treated as universal latency guarantees.
Variable length is a practical feature
Earlier audio generators often produced a fixed window even when a user wanted a short sound. Stable Audio 3 accepts duration at per-second granularity. A designer can request a brief transition, a 30-second bed, a two-minute cue, or a longer Medium composition without computing unused latent frames.
This makes both cost and memory easier to control. It does not guarantee musical closure at every arbitrary duration. A model may fade, cut a phrase, or leave an unresolved cadence. Prompt the intended structure and ending, then inspect the complete waveform.
The launch announcement highlights over six minutes for Medium and Large; the current repository expresses the maximum precisely as 380 seconds, or six minutes and twenty seconds. Small reaches 120 seconds.
Generation is only one mode
The official library supports three main inference families:
- text-to-audio from a prompt and duration;
- audio-to-audio transformation using an initial recording and noise level;
- inpainting and continuation of selected regions.
Inpainting can replace one interval or several non-contiguous intervals while preserving the rest of a track. Continuation masks the region after the source endpoint and asks for a longer output. These controls can be more efficient than regenerating an entire composition because one transition or fill is weak.
Audio-to-audio editing is powerful and risky. The user must have the right to upload and transform the source. A model license does not grant permission to edit a commercial recording, imitate an artist, or reuse protected stems.
LoRA changes customization
Stability publishes LoRA training support for the Small and Medium families. A LoRA adapts a base model with a relatively small set of trainable parameters and can be adjusted or stacked at runtime.
Use this for a licensed sound library, game environment, product sonic palette, or production genre that the general checkpoint does not represent well. Keep the dataset provenance, contributor terms, preprocessing, training configuration, base checkpoint, and LoRA version.
Avoid naming a living artist as the customization target when the real requirement is describable through musical properties. Document tempo, instrumentation, arrangement, harmonic density, production era, space, and emotional function instead.
Training data and output rights
Stability says the family was trained on fully licensed data. The official model card identifies 806,284 AudioSparx recordings and 472,618 Freesound recordings, for about 1.28 million examples. The text-conditioning language is listed as English.
Licensed training data is a meaningful provider claim, not a blanket clearance for every output. A generation can still resemble existing music, include unexpected vocal-like material, or conflict with a trademark or client exclusivity clause. Run similarity and editorial review for high-value releases.
The Stability AI Community License says qualifying users own their outputs and can commercialize them under its terms. Stability’s launch article directs organizations above $1 million in annual revenue toward an Enterprise license. Terms can change and have legal consequences, so review the current license rather than relying on a blog summary.
Hardware and reported performance
Small models run without a GPU. The repository lists CPU inference and optimized TFLite, Apple MLX, and TensorRT routes. Medium requires CUDA in the standard path and uses Flash Attention 2.
Published peak VRAM for Medium grows from about 5.07 GB for five seconds to 6.52 GB for 380 seconds on the measured setup, with chunked decoding reducing some peaks. Those are reference figures, not a complete capacity plan. Model weights, framework overhead, concurrent jobs, audio buffers, and fine-tuning add memory.
Benchmark the target machine with the chosen runtime. Record cold-start time, generation time, memory, output duration, clipping, and failure rate. An optimized H200 result cannot predict a laptop’s thermal behavior or a shared GPU’s throughput.
What to evaluate in generated audio
Music review should cover structure, melody, rhythm, harmony, transitions, stereo image, noise, clipping, and ending quality. Sound effects need onset, duration, realism, spectral balance, and fit with the visible event.
Also measure prompt adherence and diversity. If ten samples produce the same arrangement, the model may not support the variation a game or campaign needs. If every output is different, deterministic editorial iteration may be difficult.
Keep generation metadata, model and checkpoint ID, prompt, duration, seed when exposed, LoRA stack, sample rate, checksum, and reviewer decision. Separate raw generations from mastered, normalized, or edited versions.
Compare checkpoints on the same prompt suite rather than assuming that a larger parameter count wins every job. Small SFX may be a better operational choice for short effects, while Medium’s longer musical structure can justify its GPU requirement. Record subjective review beside runtime measurements so deployment decisions reflect the actual asset, not only throughput.
From an audio asset to a video deliverable
Stable Audio 3 creates audio; it does not attach that audio to a finished video or decide how it should be mixed against dialogue. A generated music bed or sound effect must first pass rights, content, loudness, and duration review.
An authorized Claude or Codex MCP client can then use Medux as an external finishing layer. Stable Audio has no native Medux integration. The Claude MCP audio-track tutorial demonstrates signed upload, asynchronous task creation, polling, and download for adding an approved track to a video.
Preserve the Stable Audio prompt and checkpoint, the generated WAV checksum, the input-video checksum, the Medux task ID, and the final artifact. If the mix is wrong, retry the finishing task or revise the approved audio; do not silently generate a different composition.
Open weights make Stable Audio 3 unusually adaptable, from offline sound effects to six-minute GPU compositions. The responsible opportunity is not merely running the model. It is building a traceable system around license, source rights, evaluation, mastering, and downstream use.