Amazon SageMaker AI now supports serverless model customization for NVIDIA Nemotron 3 models, including Nemotron 3 Nano (30B total/3B active parameters) and Nemotron 3 Super (120B total/12B active parameters). The hybrid Mamba-Transformer MoE architecture enables high throughput at lower compute cost. Three fine-tuning techniques are supported: Supervised Fine-Tuning (SFT) for labeled input-output pairs, Reinforcement Learning with Verifiable Rewards (RLVR) for tasks with objectively verifiable outcomes, and Reinforcement Learning from AI Feedback (RLAIF) for aligning tone and quality. The serverless approach eliminates GPU cluster provisioning and infrastructure management. Training data must be in JSONL format, metrics are tracked via SageMaker MLflow, and evaluation options include LLM-as-a-Judge, custom scorers, and academic benchmarks. Fine-tuned models can be deployed to SageMaker Inference endpoints or downloaded from S3.
