AWS Supports Serverless Fine-Tuning for Nemotron 3

Ars Technica AI · rss · 2026-07-10

AWS explains how to perform serverless model customization for **NVIDIA Nemotron 3** on **SageMaker AI**. ### Key Highlights - Supported models include **Nemotron 3 Nano 30B** and **Nemotron 3 Super 120B**. - These models feature a hybrid **Mamba-Transformer MoE** architecture and support up to **1M token** context. - AWS highlights their advantages in inference efficiency, throughput, and multi-agent tasks. ### Available Customization Methods SageMaker AI supports three types of fine-tuning/alignment: - **SFT**: Teaches the model new behaviors using labeled samples; - **RLVR / RFT**: Optimizes tasks like tool calling, code correctness, and format adherence using verifiable rewards; - **RLAIF**: Uses another AI model to provide feedback, reducing manual annotation costs. ### Main Selling Points - No need to manage your own GPU clusters, distributed training, and checkpoints; - Ideal for transforming general open-weight models into enterprise-specific models; - Emphasizes the strategy of fine-tuning smaller models to rival larger ones, thereby saving costs.

Original post →

More from Infra

Infra channel →