JA EN
shiichan

Nemotron 3.5 Lightning Now Supports Serverless Fine-Tuning on SageMaker AI!

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

Hi, I'm Shii-chan! Today I've got exciting news from AWS about training big AI models on your own data, so let's dig in together.

AWS What's New aws.amazon.com

What was announced?

According to AWS What's New, Amazon SageMaker AI now supports serverless model customization for NVIDIA's Nemotron 3.5 Lightning model. You can use three different techniques:

  • SFT (Supervised Fine-Tuning): improve accuracy on domain-specific tasks using labeled data
  • DPO (Direct Preference Optimization): align outputs with your organization's tone using preference data
  • RFT (Reinforcement Fine-Tuning): boost performance on new tasks using reward signals

Nemotron 3.5 Lightning is one of NVIDIA's open-weight models. It uses a hybrid Mixture-of-Experts architecture with 3 billion active parameters and 30 billion parameters in total.

The story so far

Customizing a foundation model with your own data usually meant setting up GPU clusters and orchestrating the training pipeline yourself. Just getting the infrastructure ready could be a project on its own.

What changes

With serverless model customization, SageMaker AI takes care of infrastructure provisioning and training orchestration for you, so you can focus on your data and evaluation instead. If a smaller, right-sized model trained on your own data can match frontier-model quality on your tasks, you also get to save on cost and latency.

Dive Deep

Serverless customization for Nemotron 3.5 Lightning is currently available in four regions:

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

I'm happy to see Tokyo on that list! You can kick off a customization job from the model page in SageMaker Studio, or call it programmatically through the SageMaker Python SDK.

Wrap-up

  • Nemotron 3.5 Lightning can now be customized serverlessly on SageMaker AI
  • Supports SFT, DPO, and RFT
  • The model uses a hybrid MoE architecture with 3B active / 30B total parameters
  • Available in four regions, including Tokyo
  • Great news for teams who want frontier-level quality without managing training infrastructure themselves!

Related posts