Stable Audio 3: The First Open Music Model Built to Run in Production, Not Just Demo

𝟮 𝗺𝗶𝗻𝘂𝘁𝗲𝘀 𝗼𝗳 𝗺𝘂𝘀𝗶𝗰 𝗼𝗻 𝗮 𝗠𝗮𝗰 𝗖𝗣𝗨 𝗶𝗻 𝟲 𝘀𝗲𝗰𝗼𝗻𝗱𝘀. 𝗡𝗼 𝗚𝗣𝗨.
Stability AI released Stable Audio 3 — and the interesting part isn't the audio quality.
It's where it runs.
━━━━━━━━━━━━━━
The family:
→ Small-Music / Small-SFX — 433M, CPU, up to 2 min
→ Medium — 1.4B, CUDA, up to 6m 20s, peaks at ~6.5 GB VRAM
→ Large — 2.7B, API and enterprise only, not open
Three of four ship open weights.
━━━━━━━━━━━━━━
Why it's fast:
→ A new autoencoder (SAME) compresses stereo 44.1 kHz audio 4096× into 256-dim latents at ~10.76 Hz
→ Variable-length diffusion — a 20s clip costs 20s of compute, not the max sequence length
→ 8-step ping-pong sampling after adversarial post-training
→ No classifier-free guidance at inference — CFG is baked in during distillation, so you pay one forward pass per step instead of two
Medium generates 380 seconds of audio in 1.31s on an H200.
━━━━━━━━━━━━━━
The part most people will skip:
It edits. Inpaint one region, inpaint multiple disconnected regions, continue a track, restyle an existing recording — everything outside the mask stays untouched.
Plus LoRA training in-repo. Stackable, adjustable at runtime.
That's a production engine, not a demo.
━━━━━━━━━━━━━━
The licensing, stated plainly:
→ Trained on fully licensed data
→ Community License: you own your outputs, commercialise freely
→ Above $1M revenue: Enterprise License, with legal indemnification
→ Code is MIT, weights are not
In a market where the big music generators are in court, "we can tell you where the training data came from" is a feature.
━━━━━━━━━━━━━━
Honest limits:
→ Instrumental only. No vocals.
→ Small models are domain-locked — music or SFX, pick one
→ Prompt prefixes are mandatory on most variants or output degrades
━━━━━━━━━━━━━━
Where this lands: on-device game audio, foley copilots, offline sonic branding, editors that regenerate only the selected bar.
Not one giant cloud model. Small specialised audio models fine-tuned inside the product.
We build exactly this — on-prem and edge inference, LoRA-adapted audio and voice models, deployed inside your infrastructure.

Core claims are accurate, with two small numeric slips.

• Mac CPU speed claim — directionally correct but imprecise. Stability AI's own model card states the models generate music and sounds in less than 2s on an H200 GPU and in less than a few seconds on a MacBook Pro M4, not a specific "2 minutes in 6 seconds" figure. Close in spirit, not an exact stated benchmark. Hugging Face
• Small models — 433M parameters — slightly off. The published diffusion-transformer parameter count is 459M, not 433M, for both small-music and small-sfx.
• Medium — 1.4B, CUDA, up to 6m20s — accurate, matches the paper's spec exactly.
• Large — 2.7B, enterprise/API only, not open — accurate.
• "Three of four ship open weights" — accurate (small-music, small-sfx, medium are open; large is enterprise-only).
• SAME autoencoder: 4096× compression, 256-dim latents, ~10.76 Hz — accurate, matches the technical paper precisely.
• Variable-length diffusion (a 20s clip costs 20s of compute) — accurate; inference cost scales with output duration rather than a fixed max length.
• 8-step ping-pong sampling, no CFG at inference (baked into distillation) — accurate, matches the paper's three-stage training pipeline description.
• "Medium generates 380 seconds of audio in 1.31s on an H200" — accurate, matches the paper exactly.
• Editing: single-region inpaint, multi-region inpaint, continuation, restyle — accurate; Stability's own materials describe single-segment editing, multi-segment editing, and causal continuation. "Restyle" is a reasonable gloss on audio-to-audio editing mode, though not the exact term Stability uses.
• LoRA training in-repo, stackable, adjustable at runtime — accurate. Stability's repo describes personalization through LoRA fine-tuning that adapts any model to a target style, stackable and adjustable at runtime. GitHub
• Licensing: fully licensed training data, Community License (own outputs, commercialize freely), Enterprise License above $1M revenue with indemnification — accurate on all points.
• "Code is MIT, weights are not" — the inference/training code repo is separately licensed (community tooling around it, like the LoRA trainer, is MIT), while the model weights fall under the Stability Community/Enterprise License — this distinction is correctly drawn.
• Instrumental only, no vocals — accurate, repeatedly confirmed across sources.
• Small models domain-locked (music or SFX, not both) — accurate; medium and large handle both domains, small variants are split.
• Prompt prefixes mandatory — accurate for AudioSparx-trained variants (small-music, medium, large).

One correction to make before posting: change 433M to 459M for the small models.