Customizing Amazon Nova models on SageMaker AI

You can customize Amazon Nova models, including the enhanced Amazon Nova 2.0 models, through recipes and train them on SageMaker. These recipes support techniques such as supervised fine-tuning (SFT) and Reinforcement Fine-Tuning (RFT), with both full-rank and low-rank adaptation (LoRA) options.

The end-to-end customization workflow involves stages like model training, model evaluation, and deployment for inference. This model customization approach on SageMaker provides greater flexibility and control to fine-tune its supported Amazon Nova models, optimize hyperparameters with precision, and implement techniques such as LoRA parameter-efficient fine-tuning (PEFT), full-rank SFT, RFT, and Continued Pre-Training (CPT).

Customization approaches

SageMaker offers two approaches for customizing Amazon Nova models:

UI-based experience – Use to customize Amazon Nova models through a simple, guided interface. This approach provides an end-to-end workflow including training, evaluation, and deployment without writing code. The UI-based experience is ideal for rapid experimentation, proof-of-concept development, and users who prefer a visual workflow.

Code-based experience – Use the SageMaker Python SDK, Nova SDK and training recipes to customize models programmatically. This approach offers greater flexibility, allowing you to configure advanced hyperparameters, integrate with CI/CD pipelines, and automate training workflows. The code-based experience is recommended for production workloads, complex customization requirements, and teams with established MLOps practices.

Customization platforms

AWS offers three platforms for customizing Amazon Nova models, each designed for different use cases and requirements:

Amazon Bedrock – Provides the easiest and fastest path to model customization with minimal setup. Bedrock handles all infrastructure management automatically, allowing you to focus on your data and use case. This platform is ideal when you need the quickest time-to-value and prefer a fully managed experience.

SageMaker training jobs – Provides a fully managed environment for customizing Amazon Nova models where you don't need to create or maintain any clusters. The service automatically handles all infrastructure provisioning, scaling, and resource management, allowing you to focus on configuring your training parameters and submitting your job. This platform offers a balance between ease of use and flexibility, supporting techniques like Parameter Efficient Fine-tuning (PEFT), Full rank fine tuning, and Reinforcement Fine-Tuning (RFT).

SageMaker HyperPod – Offers a specialized environment for large-scale distributed training by requiring you to create and manage EKS clusters with restricted instance groups (RIGs). This platform gives you maximum flexibility in configuring your training environment with specialized GPU instances and integrated Amazon FSx for Lustre storage, making it particularly well-suited for advanced distributed training scenarios, ongoing model development, and enterprise-scale customization workloads.