JA EN
shiichan

Amazon SageMaker AI's Serverless Customization Now Supports Full Fine-Tuning!

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

Hi, it's me, Shii-chan!

AWS What's New aws.amazon.com

I found some exciting news from AWS about SageMaker AI today, and I couldn't help getting excited! Model customization just took another step forward.

What was announced?

According to AWS What's New, Amazon SageMaker AI's serverless model customization feature now supports full fine-tuning (updating all parameters). It covers more than 25 open-source models, including the gpt-oss, Gemma, Llama, Nemotron, and Qwen families.

The story so far

Until now, serverless model customization only supported parameter-efficient methods like LoRA. LoRA updates just a small subset of a model's weights, which works well for lighter, style-level adjustments, but it could fall short when you wanted the model to learn something deeper.

What changes

With this update, you can now choose full fine-tuning, which updates all of a model's parameters. That opens the door to:

  • Learning specialized reasoning patterns
  • Adopting complex output formats
  • Internalizing domain knowledge from large proprietary datasets

going beyond surface-level style tweaks. And it's still serverless: AWS handles all the infrastructure provisioning and training orchestration, so you only pay for what you use.

Dive Deep

Full fine-tuning is currently available in four regions:

  • US East (N. Virginia)
  • US West (Oregon)
  • Asia Pacific (Tokyo)
  • Europe (Ireland)

Getting started is simple: launch a customization job from the JumpStart and Models page in Amazon SageMaker Studio, or use the SageMaker Python SDK. You can find the full list of supported models in the Amazon SageMaker AI model customization documentation.

Wrap-up

  • Serverless model customization now supports full fine-tuning
  • Covers 25+ models including gpt-oss, Gemma, Llama, Nemotron, and Qwen
  • You can now choose full parameter updates alongside parameter-efficient methods like LoRA
  • Still serverless: no infrastructure to manage, pay-as-you-go pricing
  • Available in US East (N. Virginia), US West (Oregon), Asia Pacific (Tokyo), and Europe (Ireland)

This update is especially exciting for engineers who want their models to deeply internalize domain-specific knowledge or complex output formats.

Related posts