AWS adds full fine-tuning to SageMaker serverless customization

AWS announced full fine-tuning support for SageMaker AI Serverless Model Customization in a November 18 technical post, letting developers update a model’s complete set of parameters without provisioning training clusters themselves.
The service handles compute allocation and checkpoint management. Developers specify a base model, dataset and training settings through the SageMaker Python SDK; switching between full fine-tuning and low-rank adaptation changes the training-type setting.
AWS illustrated the feature with Qwen3-4B and a 110,000-example coding dataset, divided into training, validation and held-out test sets. Its evaluation combined automated judging, text-similarity measures and general-knowledge testing. These are company experiments, not evidence that full fine-tuning always beats cheaper adaptation methods.
The post explicitly warns about catastrophic forgetting: aggressive training can improve a narrow task while damaging broader capabilities. AWS also reports cases where LoRA performed better on its automated judge metric.
The addition gives teams another managed customization option, but its practical value depends on testing task-specific gains against training cost and the capabilities a specialized model may lose.
Sources and citations
Revision history
Published .