Amazon SageMaker AI Serverless Model Customization now supports full fine-tuning, letting teams adapt every weight in a supported model without provisioning GPU clusters or configuring distributed training. The service handles compute allocation and checkpoints, then registers the result for evaluation or deployment.
Full fine-tuning can learn changes that Low-Rank Adaptation, or LoRA, cannot represent because LoRA updates only small adapter matrices. In AWS experiments, a Qwen3-4B model trained on 110,000 coding examples reached 90% semantic correctness as the dataset grew. The comparison used the same prompts, targets and evaluation code for both methods.
The stronger method is also easier to damage. Updating every parameter can overwrite broad knowledge, a problem called catastrophic forgetting. AWS found that aggressive learning rates caused sharp declines on math and general-knowledge tests. LoRA also slightly beat full fine-tuning on one automated judging setup, showing that the more expensive method is not universally better.
The practical choice depends on data volume and task complexity. Teams can switch between full and LoRA training in the SageMaker SDK, but should compare held-out task quality and general-capability benchmarks before deploying either result.