Stable Audio 3 sizes: which one for your job
Stable Audio 3.0 ships Small SFX, Small, Medium and an API-only Large. Match size to job, and see how Sume's Music Router handles length with a prompt.

Stable Audio 3.0 comes in four sizes: Small SFX and Small at 459M parameters for up to 2 minutes, Medium at 1.4B for up to 6:20, and Large at 2.7B, which is available only through the API, a partner, or enterprise licensing. Pick the smallest size that reaches your length, and use Large only if you accept the API route.
The sizes
Music Business Worldwide lists the following. It also reports an AudioSparx training set of 806,284 files, plus LoRA fine-tuning and audio inpainting.
| Size | Parameters | Maximum length | Access |
|---|---|---|---|
| Small SFX | 459M | Up to 2 min | Per report |
| Small | 459M | Up to 2 min | Per report |
| Medium | 1.4B | Up to 6:20 | Per report |
| Large | 2.7B | Not stated here | API, partner or enterprise only |
Matching size to job
- Sound effects and short stingers: Small SFX.
- Loops, intros and bumpers under two minutes: Small.
- Full songs and long beds: Medium.
- Large: only when you need it and can use the API or an enterprise arrangement.
How Sume's Music Router handles length
Sume's Music Router takes a prompt of 1 to 5000 characters and picks the engine when model is omitted or set to sume/music-auto, which is Lyria 3.5 today. Routable ids are sume/music-auto, lyria-3.5 and lyria-3-pro. The duration and duration_seconds fields are rejected, so you steer length in the prompt, for example by asking for a 2-minute track or writing timed sections like [0:00-0:30] Intro.
Every Music Router model charges the fixed Music price per audio generation. Read the audio artifact from result.artifacts[] where type is audio. Sume's docs do not list Stable Audio ids, so this is a comparison of approaches, not a drop-in swap.
What to test
Generate the same brief on both routes and compare three things: whether the length lands within your tolerance, whether a loop point is clean, and whether the result is licensed for your use. Check each provider's terms for commercial use before you ship.
Sources
Related posts
More in Models
- Tidal stopped paying wholly AI tracks: what a video score is
Tidal stopped paying royalties on wholly AI tracks on July 15, 2026. What that means for a video soundtrack, and how Sume's Music Router fits.
- Trending TikTok metadata to a Seedance 2.5 clip
Sume's trending-videos search returns TikTok metadata for $0.10 a call, not video files. Use it to write a brief, then render an original Seedance 2.5 clip.
- Veo 3.1 extend adds 7 s up to 20 times, 720p only: plan a long clip
Google's Veo docs let you extend a clip by 7 seconds up to 20 times, at 720p only. Read the arithmetic, then compare it with one-call 30 second models on Sume.
- Voxtral Mini Transcribe 2 and Realtime v26.02: what Mistral lists
Mistral lists Voxtral Mini Transcribe 2, Voxtral Realtime v26.02 and Voxtral TTS v26.03. How to prepare video audio for any transcription model.
Written by Sume