Amazon SageMaker AI offers Full Fine-Tuning of models without infrastructure management
Amazon's SageMaker AI Serverless Model Customization service now supports Full Fine-Tuning (FFT) — updating all of a model's parameters without having to deal with GPUs and clusters. According to the company, it outperforms the LoRA method on complex tasks and large datasets.
Amazon has expanded its SageMaker AI Serverless Model Customization service to support Full Fine-Tuning (FFT) — a technique that, when fine-tuning a language model, updates all of its parameters at once, in contrast to the parameter-efficient method LoRA, which adjusts only a small subset of parameters using low-rank matrices. According to Amazon, LoRA's low-rank constraint limits the expressive capacity of the resulting model and, for complex tasks and datasets exceeding tens of thousands of examples, leads to an "accuracy ceiling," where further increasing the rank no longer improves quality. According to the company, FFT bypasses this ceiling by adjusting the model's weight matrix as a whole.
FFT is available as a serverless function, so the user doesn't need to deal with GPU provisioning, cluster setup, or distributed training configuration — it's enough to specify the base model, dataset, and hyperparameters, and the service handles the rest. Amazon demonstrated this with an experiment involving the Qwen3-4B model fine-tuned using supervised fine-tuning on the Magicoder-Evol-Instruct-110K dataset (110 000 instruction pairs focused on programming, split 90/5/5 into training, validation, and test data). For a dataset of over 100 000 examples, FFT achieved, according to the company, 90.0% on the LLM-as-Judge metric for the semantic correctness of generated code, with the lead over LoRA, according to the charts presented, also showing up in the BLEU and ROUGE-2 metrics.
The source article further describes the spectral properties of weight matrices that, according to the company, explain this difference in quality, as well as hyperparameter tuning recommendations — however, these sections are not available in full in the source. See the source article for details.
Why it matters
For teams fine-tuning language models on their own data, this is another choice between the faster and cheaper LoRA method and the more powerful but more demanding FFT method — now without having to build their own GPU infrastructure for full fine-tuning. This lowers the operational barrier to experimenting with FFT for tasks where LoRA runs into quality limits, especially with large and complex datasets.
Two audiences, two different impacts
What this means
For individuals
ML engineers and developers working on model fine-tuning now have access to a serverless option for full fine-tuning without having to manage their own GPU cluster, which makes it easier to choose a method based on task complexity and dataset size.
For a business
Companies using AWS for language model customization can consider FFT instead of LoRA for tasks with large and complex datasets to achieve higher output quality, without having to invest in their own distributed training infrastructure.
DevelopmentCheck the original
Event sources
clearly official source · 1 publisher, 0 independent. We count feeds from the same owner only once.