SageMaker AI Serverless Adds Full Fine-Tuning (FFT) for Model Customization
Amazon SageMaker AI Serverless Model Customization now supports Full Fine-Tuning (FFT), enabling developers to fully tune all parameters of large language models (LLMs) serverlessly and outperform Low-Rank Adaptation (LoRA) for complex, large-scale tasks.
What changed?
Amazon SageMaker AI Serverless Model Customization now supports Full Fine-Tuning (FFT) for large language models (LLMs). Developers can fine-tune all model parameters serverlessly, without having to manage infrastructure, set up GPUs, or configure distributed training. Switching between FFT and LoRA is simplified—developers only need to change the TrainingType parameter in the SFTTrainer configuration. All benchmarking and evaluation can be done using SageMaker's built-in methods.

Why does it matter to an everyday developer?
FFT eliminates the limitations of LoRA's low-rank weight updates, enabling models to absorb richer, more complex domain-specific knowledge, particularly when working with large and diverse datasets. On tasks where datasets exceed tens of thousands of examples, FFT significantly outperforms LoRA in metrics such as semantic correctness (LLM-as-Judge), BLEU, and ROUGE-2. This brings higher quality results on complex or code-centric customization tasks. Serverless delivery lowers operational overhead, letting developers focus on model quality with no infrastructure management.
What can the developer do now?
Developers can use SageMaker AI Serverless Model Customization to launch FFT jobs by simply specifying TrainingType.FULL for SFTTrainer. They can benchmark FFT against LoRA by changing a single parameter, with all other code and evaluation pipelines remaining unchanged. Built-in evaluations like LLM-as-Judge, PRIMECODE (for BLEU, ROUGE, F1), and MMLU are available for assessing customized model performance. This enables rapid iteration in production scenarios that require the higher accuracy and expressiveness FFT provides, particularly for large, complex datasets or tasks.
