Stability AI releases Stable Diffusion 3.5

Stable Diffusion is the name of a family of text-to-image models: systems that take a written prompt and produce a picture. The 3.5 release arrived in stages. On 22 October 2024, Stability AI published Stable Diffusion 3.5 Large, a model with 8.1 billion parameters, together with Stable Diffusion 3.5 Large Turbo^[1]. On 29 October 2024 the company added a third model, Stable Diffusion 3.5 Medium, with 2.5 billion parameters^[1].

Three models, two sizes

Parameters are the learned numbers inside a network; broadly, more of them means more capacity and more computation per image. Large and Large Turbo sit at 8.1 billion parameters, Medium at 2.5 billion^[1]. The two large models differ in how they are trained rather than in size. Large Turbo is a distilled model: distillation trains one model to reproduce the outputs of another, usually to make sampling cheaper, and here it lets the model generate an image in four steps^[1]. A conventional diffusion model needs many more sampling steps, so the four-step figure is the headline difference between the two 8.1-billion-parameter variants.

License

All of the models are released under the Stability AI Community License. Under that license, commercial use is free for organisations whose annual revenue is below $1 million^[1]. Organisations at or above that threshold must contact Stability AI to arrange an enterprise license^[1]. In practice this creates two tiers of users: individuals and smaller companies can adopt the models commercially without negotiating, while larger companies have to reach a separate agreement.

Where the weights are

The model weights were published on Hugging Face^[1]. Publishing weights means the files needed to run the model can be downloaded rather than only accessed through someone else's servers. Alongside that, Stability arranged hosted access through the Stability API, Replicate, Fireworks AI, DeepInfra and ComfyUI^[1] — services that run the model for users who do not want to manage the hardware themselves.

Hardware requirements

Stability states that Stable Diffusion 3.5 Medium needs 9.9 GB of VRAM, excluding text encoders, to unlock its full performance^[1]. VRAM is the memory on a graphics card, and it is usually the binding constraint when running an image model locally: if the model and its intermediate computations do not fit, it either fails to run or has to be run in a reduced form. The stated figure gives prospective users a concrete bar to check their own hardware against, and the exclusion of text encoders signals that the total footprint of a working setup is larger than that single number.

Why the release matters

The release put several options on the table at once: a large model, a fast distilled variant of it, and a smaller model that followed a week later^[1]. The licensing split at $1 million in annual revenue, combined with weights on Hugging Face and a set of hosted providers, meant the same models could be run locally by individuals, rented by the image through third-party services, or licensed separately by larger companies^[1]. For people following open image models, the practical questions around a release of this kind are usually size, sampling speed, license terms and memory footprint; Stability published a figure for each of them on the day, which makes the models easy to compare against what a user already has.