Stable Diffusion 3.5 Large lands on AWS SageMaker

Deploying open-source image generators inside an enterprise environment is usually an infrastructure headache. Between data privacy rules, spinning up GPUs, and managing private clusters, most teams get stuck before generating a single picture.

Stability AI and AWS just teamed up to bypass that hassle. Stable Diffusion 3.5 Large, the company's flagship 8.1B parameter text-to-image model, is now directly available inside Amazon SageMaker JumpStart.

The first major SD upgrade on AWS in over a year

Believe it or not, this marks the first new Stable Diffusion release on SageMaker JumpStart since SDXL 1.0 dropped back in July 2023. A lot has changed in generative media since then, and Stability AI needed to get its latest weights in front of cloud customers.

Interestingly, AWS revealed that SD 3.5 Large was actually trained on Amazon SageMaker HyperPod. The model specializes in 1-megapixel outputs across 3D, photorealistic, line art, and painterly styles, with major upgrades in multi-subject generation and prompt adherence.

ML teams can deploy the model with just a few clicks inside SageMaker Studio or via the SageMaker Python SDK. Crucially, it runs within a customer's Virtual Private Cloud (VPC), ensuring complete data isolation and network security.

The deployment is currently live across global AWS regions, including US East, US West, Europe, Asia Pacific (including Hong Kong and Osaka), and Beijing.

Why it matters

We think this partnership is a pragmatic play for Stability AI's enterprise survival. Facing intense competition from Midjourney on the consumer side and Black Forest Labs' Flux in the open weights space, Stability AI needs enterprise distribution where companies pay real cloud bills.

For enterprise teams in gaming, advertising, and retail, this eliminates the compliance barrier. Companies that cannot send sensitive concept art or customer data over external APIs can now run top-tier image generation inside their own perimeter with native MLOps tooling like SageMaker Pipelines.

The catch, however, is infrastructure cost. Hosting an 8.1B parameter model on dedicated cloud instances is far from cheap. Unless you have high-volume, continuous generation needs, running persistent instances on SageMaker will quickly rack up serious cloud spend compared to pay-per-call APIs.