Stability AI releases Stable Diffusion 3.5

Stability AI released Stable Diffusion 3.5 in three variants: Large (8.1 billion parameters) for professional-quality output at up to one megapixel, a distilled Large Turbo capable of comparable image quality in four generation steps, and a Medium model (2.5 billion parameters) intended to run on consumer hardware. Medium followed a week after the Large variants.

The models were released under Stability’s Community Licence — free for non-commercial use and for commercial use up to $1 million in annual revenue, with paid licensing above that — and made available for download via Hugging Face and GitHub as well as through Stability’s own API and third-party platforms including Replicate and DeepInfra.

The release came as Stability worked to recover ground against Black Forest Labs’ Flux and Midjourney, both of which had drawn users away from Stable Diffusion on image quality and prompt adherence following the company’s own leadership turmoil and the comparatively poor reception of Stable Diffusion 3’s initial release earlier in 2024. Positioning 3.5 around accessibility — permissive licensing, a range of model sizes, and openly downloadable weights — was Stability’s answer to that: compete on openness and hardware flexibility where it could no longer reliably compete on raw output quality alone.