Stable Audio 3: Local Setup, License, Pricing and ComfyUI

👁 113 reads

Stable Audio is an AI music and sound effect generator developed by Stability AI. The latest Stable Audio 3 model family (May 2026) produces instrumental tracks and audio effects up to 6 minutes and 20 seconds long, though it does not sing lyrics. Three of the four models offer open weights, meaning you can run them for free on your own hardware-including for commercial projects if your business earns under $1M in annual revenue.

What is Stable Audio?

Stable Audio is a dedicated suite of audio models created by Stability AI, the creators of Stable Diffusion. The original version premiered in September 2023, followed by version 2.0 in April 2024 and version 2.5 on September 10, 2025. Stability also previously released open-weight variants known as Stable Audio Open (up to 47s) and Open Small (11s).

Stable Audio 3 marks a significant leap forward. The Small variant creates audio up to 2 minutes, while Medium and Large reach 380 seconds (6 minutes 20 seconds). All versions output 44.1 kHz stereo sound. Thanks to an optimized diffusion training stage, generation requires just a handful of steps rather than dozens - the official ComfyUI workflow uses only 8 steps.

One essential limitation: Stable Audio does not produce vocal singing. Official documentation explicitly clarifies that these models cannot deliver coherent spoken or sung words, only ambient “vocal textures” without semantic phrasing. Prompts should be written in English because the training data was English-centric; non-English prompts yield degraded performance. If you need full songs with intelligible lyrics, turn to Suno, Eleven Music, or Lyria.

Stable Audio 3 model variants

These figures come directly from Stability AI’s GitHub repository. Note that the Stability blog cites 459M parameters for the Small edition, while one repository document specifies “around 4.75 minutes” instead of 380 seconds for Medium.

All variants provide three core generation modes:

• Text-to-audio - Natural language description paired with a duration parameter in seconds.
• Audio-to-audio - Uses an existing audio file as an acoustic starting point, where an init_noise_level parameter dictates how closely the model adheres to the source material.
• Inpainting and audio continuation - Replaces designated segments (including multiple non-contiguous regions) or seamlessly extends an existing track.

Additionally, the architecture supports LoRA fine-tuning, allowing you to train the model on your own audio library to replicate custom sonic aesthetics. Stability released official documentation for this workflow for the first time. LoRAs are trained on “base” checkpoints and applied on top of standard release weights. On fal.ai, an automated trainer is available (fal-ai/stable-audio-3-trainer), charging $3 per 1,000 steps.

Artificial Analysis benchmark rankings

In the blinded AA-Music-Instrumental v1.1 evaluation (recorded around October 7, 2026, across 1,000 prompts spanning 17 genres), the competitive standings were as follows:

Stable Audio 3 Medium scores roughly on par with MiniMax Music 3.0, with the critical advantage that you can download Medium and host it on your own rig. However, it trails Suno v6 by more than 140 Elo points. Because Stable Audio lacks singing capabilities, it does not compete on vocal leaderboards. For full comparative data across all systems, check our AI music generator rankings.

Commercial licensing and usage rights

The open models fall under the Stability AI Community License (dated July 5, 2024). Reviewing the legal terms highlights several critical requirements:

• Under $1M in annual revenue: You can deploy the models commercially without paying license fees. This threshold accounts for total aggregate revenue across all affiliated entities.
• Over $1M in annual revenue: You must negotiate an Enterprise license with Stability AI, which includes legal indemnification protections.
• Ownership of output: You own the generated audio files “to the extent permitted by applicable law.”
• Registration: Commercial operators must complete an official registration on Stability AI’s platform.
• Attribution: Any product, feature, or service built directly on top of the model must display “Powered by Stability AI” in its documentation, website, or service description.
• Restrictions: You may not use the model weights or generated audio outputs to train competing generative models.

The Hugging Face repository is gated: downloading requires submitting your name, email, country, and intended operational use. The Medium checkpoint incorporates a text encoder derived from the Gemma family, thereby binding users to the supplemental Gemma Terms of Use. On the training side, Stability utilized licensed music from the AudioSparx catalog along with CC0 audio assets from Freesound - totaling roughly 1.28 million recordings per the research paper. This clear provenance offers safety for enterprise and advertising clients, especially as platforms like Suno and Udio have faced copyright litigation from major recording labels over unverified training data.

Stable Audio pricing breakdown

Running the models locally incurs only electricity costs. If you prefer cloud workflows, options include the web platform, the official Stability API, and fal.ai.

Pricing verified as of October 8, 2026. Track generation estimates apply to single mixed master files; multi-track stem generation (currently experimental) consumes roughly one-third more credits. Unused credits expire monthly and do not roll over. All paid tiers include digital audio workstation (DAW) plugins compatible with Ableton, Logic Pro, and Pro Tools. Note that older tiers cited in historical reviews ($11.99 Pro or $29.99 Studio) have been discontinued. Avoid unofficial domains like stableaudio3.com, stableaudio3.org, or stableaudio.net, which have no connection to Stability AI.

Developer and API pricing:

• Stability API: Stable Audio 3 costs 26 credits per generation (1 credit = $0.01, translating to $0.26 per call) for durations between 1 and 380 seconds. Stable Audio 2.5 costs 20 credits. New accounts receive 25 complimentary credits upon registration.
• fal.ai Medium: $0.0376 per generated audio clip (text-to-audio); inpainting workflows cost $0.0442 per call.
• fal.ai Large: $0.00125 per second of compute execution time.

How to run Stable Audio 3 locally in ComfyUI

The most accessible path for Windows users is running ComfyUI with the official template. Stability’s native repository implementation requires Flash Attention 2 for the Medium checkpoint and is heavily tailored toward Linux environments.

1. Update ComfyUI to its latest release (Stable Audio 3 workflows ship inside the Comfy-Org/workflow_templates package).
2. Head to the Comfy-Org/stable-audio-3 Hugging Face repository and download stable_audio_3_medium.safetensors (9.22 GB) into your models/checkpoints folder. For resource-constrained rigs, download stable_audio_3_small_music.safetensors (2.27 GB).
3. Download the text encoder t5gemma_b_b_ul2.safetensors (1.18 GB) and place it inside models/text_encoders.
4. Optional: Download qwen3.5_2b_bf16.safetensors from Comfy-Org/Qwen3.5. The default ComfyUI workflow uses this compact LLM to expand concise user inputs into descriptive musical prompts across Music, Instrument, SFX, and One-shot modes.
5. Open ComfyUI, navigate to Templates, open the audio directory, and select “Stable Audio 3 Medium” (workflow file audio_stable_audio_3_medium).
6. Enter an English prompt, define the track duration in seconds (150 seconds by default), and click Run. Exported MP3 renders appear in output/audio/stable_audio_3.

Hardware requirements: According to Stability, Medium demands between 5.07 GB VRAM (for 5 seconds of audio) and 6.52 GB VRAM (for 380 seconds), while chunked decoding methods reduce consumption by over 1 GB. The Small edition can run purely on CPU, including Apple Silicon Macs. However, because the packaged Medium checkpoint file for ComfyUI is 9.22 GB, real-world memory allocation can run higher, as demonstrated in our hardware tests below.

Hands-on testing and performance benchmarks

On October 3, 2026, running Stable Audio 3 Medium via fal.ai produced a 60-second backing track in 18.7 seconds. Direct audio clips and side-by-side evaluations against singing models can be found in our AI music generator guide.

Local benchmark on RTX 4090 (October 8, 2026)

I set up Stable Audio 3 Medium inside ComfyUI using stable_audio_3_medium.safetensors (9.22 GB) and the t5gemma encoder (1.19 GB). Generation parameters adhered to the official preset: 8 steps, CFG 1.0, and LCM sampler. Prompts were written manually in English without relying on the automated LLM expansion node.

• Inference speed is remarkable: Once loaded into VRAM, generating 60 seconds of complete stereo audio takes just 4 seconds. The initial cold run suffered latency because the heavy weights were loaded over a USB drive transfer bottleneck (~44 MB/s).
• VRAM consumption: ComfyUI peaked between 7.9 GB and 11.8 GB VRAM. While higher than the 6.5 GB figure cited in Stability’s base repository, it still fits comfortably within any standard 12 GB consumer GPU.
• Automated AI judge scoring (Gemini 3.8 Flash and Qwen 3.8 Omni Flash via OpenRouter across 3 audio samples): Sound quality averaged 4.0/5, style adherence scored 4.6/5, and musicality stood at 4.0/5. Qwen noted slightly boxy percussion and repetitive guitar riffs, whereas Gemini praised the opening indie pop instrumental as production-ready library quality.
• Vocals were completely absent across all renders, precisely aligning with model specifications.

In our wider benchmark battery, the same stylistic aesthetic was evaluated with full vocals against Eleven Music, Lyria 3 Pro, ACE-Step, and MiniMax Music - full comparisons are available in our AI music generator leaderboard.

Best practical applications

• Background audio for videos and reels - Easily match exact video timelines by setting precise durations, tempos (BPM), and harmonic moods.
• Sound effects and foley - Small SFX and Medium excel at cinematic impacts, risers, mechanical textures, and room ambiances. Use brief target durations and include tags like TrackType: SFX.
• Loops and instrument samples - Generate isolated instrument tracks (TrackType: Instrument), synth hooks, drum loops, or sub-bass lines for modern music production.
• Audio redesign and stem modification - Leverage audio-to-audio to resynthesize rough sketches into fresh genres, or use inpainting to regenerate transition sections.

Stability’s documentation emphasizes that structuring prompts with AudioSparx metadata tags produces the most predictable outputs: e.g., TrackType: Music, VocalType: Instrumental, Genre: Funk, Instruments: Guitar, Saxophone, supplemented by an explicit tempo in BPM. For broader audio tooling, browse our curated roundup of AI tools for music and audio production.

Competitive landscape: Stable Audio vs. competitors

If you need an open-source model that actually handles vocals, ACE-Step 1.5 (MIT licensed) is capable of synthetic singing - during our October 8 testing with non-English lyrics (tested in Polish), it accurately enunciated 101 out of 104 words. For full songs with vocals, Suno, Eleven Music, or MiniMax Music 3 remain superior. For copyright-safe instrumental production and local deployment, Stable Audio 3 stands out. Learn more in our dedicated reviews for Suno, Eleven Music, Lyria, Mureka, and MiniMax Music.

Who should use Stable Audio 3?

• Video editors and content creators needing clean, risk-free instrumental backgrounds matched to precise clip lengths without royalty claim anxieties.
• Home lab enthusiasts and producers with NVIDIA GPUs looking for unlimited, subscription-free music generation - explore our guide to running local AI models.
• Game developers and UI sound designers who need fast, procedural foley and interface audio effects.
• Not recommended if your primary objective is generating fully vocalized songs.
• Not recommended if you only care about maximum instrumental Elo quality, where Suno v6 remains the front-runner.

Frequently asked questions

What is Stable Audio?

Stable Audio is an AI music and sound effect generation engine built by Stability AI. Using natural language prompts, it synthesizes high-fidelity instrumental music, ambient loops, and foley sound effects. Stable Audio 3 debuted in May 2026 across stableaudio.com, official APIs, and open-weight model checkpoints.

How do I run Stable Audio 3 locally?

The easiest method is using ComfyUI: download the Medium model checkpoint (9.22 GB) and the t5gemma text encoder (1.18 GB) from the Comfy-Org/stable-audio-3 Hugging Face repository, open the default “Stable Audio 3 Medium” workflow template, and write your prompt. The Small Music model can even run on basic CPUs.

Is Stable Audio free to use?

Yes, via two distinct paths. The Free tier on stableaudio.com grants 50 monthly credits (approx. 12 tracks) strictly for personal, non-commercial experimentation. Alternatively, the open-weight Small and Medium models can be hosted locally for free, with commercial rights permitted up to $1M in annual organizational revenue.

Can I monetize music made with Stable Audio?

Yes. On the cloud platform, commercial rights require a paid tier starting at Solo ($12/month). For self-hosted open-weight models, commercial usage is covered under the Stability AI Community License for companies earning under $1M annually. Entities crossing that threshold must secure an Enterprise license. Web Free tier outputs are strictly non-commercial.

Can Stable Audio sing vocals?

No. Stable Audio 3 cannot generate intelligible sung or spoken lyrics; at most, it produces occasional background vocal textures without decipherable words. If you require songs with lyrics, choose Suno, Eleven Music, or Lyria.

Stable Audio vs. Suno: which is better?

For complete vocal tracks and bleeding-edge instrumental quality, Suno leads comfortably (1140 Elo vs. 1017 for Stable Audio 3 Large). For local offline execution, granular sound effect design, region inpainting, and custom LoRA training, Stable Audio is the superior choice.

What is the best AI instrumental generator?

In terms of pure acoustic quality, Suno v6 holds the highest benchmark ranking in instrumental mode. For cost-effective volume generation, Stable Audio 3 Medium is completely free when run locally and costs only $0.0376 per track on fal.ai. Explore the head-to-head breakdown in our AI music generator guide.

Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →

Sources and verification dates

• Stability AI: Stable Audio 3.0 Announcement (May 20, 2026).
• GitHub Stability-AI/stable-audio-3: Model architecture, VRAM usage metrics, prompting syntax, and dataset specifications - verified October 8, 2026.
• Stable Audio 3 Technical Report (arXiv, May 18, 2026).
• Hugging Face: stable-audio-3-medium and Comfy-Org/stable-audio-3 - verified October 8, 2026.
• Stability AI Community License Agreement (dated July 5, 2024).
• stableaudio.com pricing and Stability API platform pricing - verified October 8, 2026.
• fal.ai: Stable Audio 3 Medium and LoRA Trainer - verified October 8, 2026.
• Artificial Analysis: Music Instrumental Leaderboard (v1.1, approx. October 7, 2026).
• Stability AI: Stable Audio 2.5 Announcement (September 10, 2025).
• In-house hardware benchmarks conducted October 3, 2026 via fal.ai - methodology detailed in our AI music generator guide.

Kinetyra