AWS adds full fine-tuning to SageMaker serverless model customization

Amazon Web Services has extended SageMaker AI Serverless Model Customization with Full Fine-Tuning (FFT), which updates every parameter of a language model without customers having to provision GPUs or configure distributed training. The company says the method pulls ahead of the lighter LoRA approach as training datasets grow.
What AWS has introduced
Amazon Web Services made the announcement on November 18, 2026. Under the update, customers can train a model on their own data to reach higher quality while the service takes care of the infrastructure, so no GPU provisioning, cluster setup or distributed-training configuration is needed.
Fine-tuning, as the company describes it, means continuing to train a pretrained large language model on an organisation's own data so the model adapts to a particular domain, task or style instead of depending on general knowledge alone. AWS states that a fine-tuned model can perform measurably better on domain work and cost less to run, since a smaller specialised model can match or exceed a frontier model on the task at a fraction of the per-token inference price, and those savings add up across millions of requests.
How full fine-tuning differs from LoRA
Parameter-efficient methods such as Low-Rank Adaptation (LoRA) have become a practical route to fine-tuning language models, because they update only a small share of the parameters and therefore lower GPU requirements. LoRA keeps the pretrained weights fixed and inserts small pairs of low-rank matrices into the model's layers. That approach handles simple adaptations well, but its low-rank structure constrains how much the tuned model can express.
AWS describes an accuracy plateau: on complex tasks and datasets of tens of thousands of examples, raising the LoRA rank, and with it the number of trainable parameters, fails to close the gap with standard fine-tuning. The company argues the shortfall is not simply the result of fewer parameters, since a low-rank update cannot capture the full complexity of weight changes that deep task adaptation demands.
FFT instead updates the model's parameters simultaneously across its full weight matrices. According to AWS, this produces stronger convergence, better alignment of activations across layers, and a model that retains both the target task distribution and the breadth of its pretrained knowledge.
Reported results and caveats
In a coding experiment, AWS reports that FFT's advantage grows with dataset scale. With more than 100,000 training examples, FFT reached 90.0 percent on LLM-as-Judge semantic correctness across code generation tasks. In the company's radar chart comparison, the base model scored low on every measure, LoRA improved in a balanced way, and FFT led on every dimension, most markedly on BLEU and ROUGE-2 scores.
AWS also cautions that aggressive training settings caused severe losses of general knowledge. These figures come from AWS's own experiment and blog post and have not been independently verified.
Open questions
- How large is FFT's accuracy advantage over LoRA on tasks other than code generation?
- How much general knowledge is lost at which training settings, and how can customers avoid that loss?