Stability AI Releases Stable Audio 3 Medium: 2B-Parameter Audio Generation Model with 180-Second Output in Under 2 Secon
Stability AI has released Stable Audio 3 Medium, a 2 billion parameter latent diffusion model capable of generating variable-length audio up to 380 seconds. The model generates music and sound effects in less than 2 seconds on an H200 GPU, trained on 1.28 million licensed and Creative Commons audio recordings.
Stability AI Releases Stable Audio 3 Medium: 2B-Parameter Audio Generation Model
Stability AI has released Stable Audio 3 Medium, a 2 billion parameter latent diffusion model that generates music and sound effects in variable lengths up to 380 seconds (6+ minutes). According to Stability AI, the model produces audio in under 2 seconds on an H200 GPU and "a few seconds" on a MacBook Pro M4.
Stable Audio 3 is the medium version in a three-tier family (small, medium, large) of fast latent diffusion models designed for consumer-grade hardware deployment.
Technical Architecture
The model operates on a novel semantic-acoustic autoencoder that compresses audio into a compact latent space, enabling efficient generation while preserving audio fidelity. Stability AI claims the architecture encourages semantic structure in the latent representation.
The model underwent adversarial post-training to reduce inference steps while improving generation quality and prompt adherence. It requires 8 diffusion steps at inference time using a "pingpong" sampler, with a CFG scale of 1.0.
Text conditioning uses Google's pre-trained T5Gemma model (t5gemma-b-b-ul2), which is redistributed under separate Gemma Terms of Use.
Training Data
The model was trained on 1,278,902 audio recordings:
- 806,284 recordings licensed from AudioSparx
- 472,618 recordings from Freesound (266,324 CC-0, 194,840 CC-BY, 11,454 CC-Sampling+)
Stability AI reports that music recordings in the Freesound portion were identified using PANNs tagging and sent to a content detection company to verify the absence of copyrighted material. All identified copyrighted content was removed.
Key Capabilities
The model supports:
- Variable-length audio generation (up to 380+ seconds demonstrated)
- Audio inpainting for targeted editing
- Continuation of short recordings
- BPM-specific music generation
- Style and mood control through text prompts
Availability and Licensing
Stable Audio 3 Medium is available on Hugging Face under the Stability AI Community License. Commercial use requires a separate license from Stability AI. The model requires users to accept both the Stability AI license and Gemma Terms of Use, including use restrictions in Section 3.2.
Inference code is available through two libraries: the stable-audio-3 inference library and the stable-audio-tools research library. The model weights are distributed in FP32 format.
What This Means
Stable Audio 3 Medium represents a significant step in accessible audio generation, with claimed sub-2-second generation times that could enable real-time workflows for sound design and music production. The 2B parameter size positions it as deployable on consumer hardware, though actual performance will depend on available GPU memory and compute. The variable-length generation capability addresses a key limitation of fixed-length audio models, reducing computational waste for short sound effects. However, commercial users should note the dual licensing requirement and review Section 3.2 restrictions in the Gemma terms before deployment.
Related Articles
TII's Falcon-Emirati-7B scores 84.83% on Alyah, a new Emirati-dialect Arabic benchmark
The Technology Innovation Institute (TII) released Falcon-Emirati-7B, a 7B-parameter model specialized for Emirati Arabic and built on Falcon-H1-Arabic. TII claims it scores 84.83% on the Alyah benchmark, ahead of every Arabic and multilingual model it compared against.
Reflection AI unveils Beam: 501B-parameter open-weight MoE with 1M-token context
Reflection AI has unveiled Beam, a text-only mixture-of-experts model with 501 billion total parameters, 23 billion active, and a 1 million token context window. The company claims it matches Z.ai's GLM-5.2 on advanced reasoning benchmarks while using 3-4x less inference compute. Weights and the full technical report are due later this month.
Reka AI releases Rho-1, a 19B-parameter omni-model for text, image, video and robot control
Reka AI has released a research preview of Rho-1, a 19-billion-parameter omni-model that processes and generates text, images, video, and robot control actions in a single network. Reka says it uses no tool calls or external models. Context window, pricing, and benchmark scores have not been disclosed.
China Telecom's Xing4.0-29B-A4B: 29B MoE, 4B Active, 256K Context, Trained Fully on Ascend NPUs
China Telecom AI's Xing4.0-29B-A4B (formerly the TeleChat line) is a mixture-of-experts model with 29B total and 4B active parameters and a native 256K context window, extensible to 512K. The company claims it is the first model of this scale trained entirely on Ascend NPUs with MindSpore. Community GGUF quantizations from Venastine-Research are already available.
Comments
Loading...