Stable Audio 3.0 Medium

Model design and access

Adversarial post-training reduces the number of sampling steps. The card's example uses eight steps. Text conditioning uses T5Gemma, while the diffusion model operates on compressed audio. The downloadable package should be assessed as a full pipeline, including the text encoder and decoder.Stable Audio 3.0 Medium official model documentation

The separate Medium Base checkpoint is the pretrained starting point for adaptation. The post-trained Medium checkpoint is the direct generation release. Optimized variants belong to this same model family.Stable Audio 3: official inference repository

Practical use

Practical guidance. Medium is a practical choice when a team needs to keep inference and adaptation within its own environment. Measure memory with the intended track length and decoder settings. Short preview clips can help assess a prompt before requesting a full track; a small text change can still change the composition.