What changed
Amazon SageMaker AI Serverless Model Customization now supports full fine-tuning without customer-managed GPU provisioning, cluster setup or distributed-training configuration, alongside built-in evaluation and Python SDK control.
Why it matters
Teams can test full-parameter adaptation without operating training infrastructure, but the cost and quality decision still depends on data scale, base-model support and held-out task evidence rather than the convenience of the control plane.
What to do
Run LoRA and full fine-tuning on the same versioned dataset and base model, reserve a blind task holdout, and compare quality, training cost, inference cost and rollback before selecting the method.
Caveat
AWS benchmark and spectral-analysis results are vendor evidence. Full fine-tuning needs sufficient high-quality data, supported models and careful leakage, memorization and safety evaluation.