Stability AI launches Stability Audio 3.0 with music generation of up to six minutes

Stability AI, known as Stable Diffusion, has just launched Stability Audio 3.0, a family of audio models that breaks the barrier of six minutes of continuous musical generation. The new system doubles the capacity of the previous version (Stable Audio 2.0) and is already available both in web version and in reduced models for local devices.

What changes with Stability Audio 3.0?

The main novelty is the maximum generation duration: six minutes of audio with professional quality. In practice, creators can now produce complete tracks, jingles or long soundtracks without the need to edit multiple outputs. The model also supports reference audio conditioning, allowing the user to provide a musical stretch to inspire style. In addition, Stability AI trained smaller models — called "Stable Audio 3.0 Light" — that run on consumption GPUs, democratizing access.

How does Stability Audio 3.0 generate long-term audio?

According to the original TechCrunch announcement, Stability Audio 3.0 uses a latent audio-specific diffusion architecture, optimized for temporal coherence at long scales. The model is trained in millions of licensed tracks and understands descriptive textual prompts (e.g. "melancholic acoustic guitar with slow and climate rhythm"). To ensure consistency, the generator maintains a musical narrative structure, avoiding repetitions or abrupt breaks.

Why does Stability Audio 3.0 matter to Brazilian creators?

Music producers, content creators and independent game developers in Brazil now have access to an AI tool that can accelerate the prototyping of complete soundtracks. While giants like Google launch multimodal models such as Gemini 3.5 Flash, Stability AI focuses on vertical niche — and prompt support in English (with the possibility of adaptation to Portuguese) makes the tool viable for home studios. The Light version in particular eliminates cloud dependency for local generation, crucial in regions with limited connectivity.

Which versions of Stability Audio 3.0 are available?

Stability AI offers Stability Audio 3.0 in two modalities: the full version by the web API (with maximum capacity and premium quality) and the "Light" models (available for download on Hugging Face). Lightweight models can generate up to three minutes of audio with reduced quality, but run in GPUs with 8 GB of VRAM. Prices have not yet been released, but the previous version charged for generation credits — the new model should follow a similar model.

Frequently Asked Questions about Stability Audio 3.0

Is Stability Audio 3.0 free?

Not at the moment. Stability AI adopts a credit billing model for the web API and offers free light versions for non-commercial use via Hugging Face.

Is it possible to generate music with vocals?

Yes, the model accepts prompts that describe vocals (such as "female voice soul" or "male rap aggressive") and is able to synthesize sung lyrics, although the quality depends on the complexity of the prompt.

Does the model understand commands in Portuguese?

The Stability Audio 3.0 was mostly trained in English, but responds reasonably well to descriptions in Portuguese when the musical terms are universal. For best results, it is recommended to use English prompts.