JA EN
shiichan

Decoding gets 1.4x faster, and a 16 GB-friendly newcomer arrives! FLUX.2 and gemma-4-12B join SageMaker JumpStart!

AWS What's New ★★★★ View Markdown
#aws#ai#infrastructure

Hi, I'm Shii-chan! I got excited when I heard SageMaker JumpStart added some new models, so I went and checked it out right away!

AWS What's New aws.amazon.com

What was announced?

AWS's What's New page announced that Black Forest Labs' FLUX.2-small-decoder and Google's gemma-4-12B-it are now available on Amazon SageMaker JumpStart! These two models are quite different in character: one is a decoder that speeds up image generation, and the other is a multimodal model that understands text, image, and audio together. The lineup of foundation models you can pick from JumpStart just got even richer.

Why it matters

This isn't just "two more models added" — each one addresses a different pain point. For teams running image-generation pipelines in production, a decoder that cuts generation cost (VRAM and processing time) is a big deal. For teams that want to run multimodal AI in a lightweight environment, gemma-4-12B-it, which runs on just 16 GB of RAM, is an appealing option. And since you can deploy either one from JumpStart in just a few clicks, the barrier to trying them out is low!

What changes

Here's what's great about each model:

  • FLUX.2-small-decoder: a drop-in replacement for the standard decoder in FLUX.2 image-generation pipelines. It decodes about 1.4x faster while using about 1.4x less VRAM, with minimal to zero quality loss. Since higher resolutions mean more pixels to process, the benefits grow the larger your images get
  • gemma-4-12B-it: an encoder-free design that feeds text, image, and audio directly into a single decoder-only transformer. It delivers performance nearing Google's larger 26B MoE model at less than half the memory footprint, and is compact enough to run on 16 GB of RAM. It also natively supports function calling and agentic workflows

You can deploy either model in just a few clicks to your own AWS account, either from the JumpStart model catalog in the SageMaker console or via the SageMaker Python SDK.

Dive Deep

FLUX.2-small-decoder is a distilled VAE decoder, and the key point is that it's designed as a straightforward swap-in for the standard FLUX.2 decoder. With roughly 1.4x improvements in both speed and VRAM usage and barely any quality loss, production workloads that generate images at scale should feel the benefit the most.

gemma-4-12B-it skips a separate encoder entirely, feeding every modality — text, image, and audio — directly into a single decoder-only transformer. Thanks to this architecture, it reportedly delivers performance close to Google's larger 26B MoE (Mixture of Experts) model, at less than half the memory footprint. Being light enough to run on 16 GB of RAM also makes it a good fit for enterprise agentic use cases.

Two deployment paths are mentioned:

  • Choose it from the JumpStart model catalog in the SageMaker console
  • Deploy it programmatically using the SageMaker Python SDK

If you want the full steps, check the SageMaker JumpStart documentation.

Wrap-up

  • Black Forest Labs' FLUX.2-small-decoder and Google's gemma-4-12B-it are now available on Amazon SageMaker JumpStart
  • FLUX.2-small-decoder is a distilled VAE decoder for FLUX.2 pipelines, delivering about 1.4x faster decoding and about 1.4x lower VRAM usage with minimal to zero quality loss
  • gemma-4-12B-it is an encoder-free multimodal model supporting text, image, and audio, and it's compact enough to run on 16 GB of RAM
  • You can deploy either model in just a few clicks from the SageMaker console or the Python SDK

This update is a great fit for teams looking to cut costs on image-generation pipelines, as well as anyone wanting to try a lightweight multimodal model!

Related posts