AWS has added full fine-tuning to SageMaker AI Serverless Model Customization, letting teams update all a model’s weights without provisioning or managing GPU infrastructure. In AWS’s coding-task experiments, full fine-tuning scored 90% for semantic correctness at training-data scale above 100,000 examples, though that result is specific to the company’s setup and evaluation.
AWS AI Watch analysis
What happened
Full fine-tuning changes all the model’s weights; LoRA updates a smaller set of added parameters. AWS says SageMaker’s serverless service now supports the full-weight option, handling compute allocation and checkpoint management for users.
AWS compared the methods using Qwen3-4B and a dataset of 110,000 coding instruction-response pairs. It reports that full fine-tuning reached 90% on its LLM-as-judge semantic-correctness measure at 100,000-plus training examples. The company also tested model quality with code-generation and general-knowledge benchmarks. Read AWS’s account of the experiments and setup.
Why it matters
LoRA is popular partly because it is less demanding to train. AWS’s results make the case for another option when teams have a large, high-quality dataset and need deeper task adaptation: they can try full fine-tuning without first assembling a GPU cluster.
That is a useful reduction in setup friction, not a guarantee of better models for every job. The reported score comes from AWS’s own coding experiments, and the right method still depends on the task, data and evaluation criteria. Full fine-tuning also changes more of the model, so checking for lost general capabilities matters.
Our read
This is a meaningful addition to the model-customisation toolkit, especially for teams that have been put off full fine-tuning by infrastructure work. AWS has supplied a concrete experiment rather than just a new button, but the 90% figure is a result to investigate, not a universal benchmark verdict. Try both approaches on representative data and keep a test for what the model must not forget. The GPU-cluster wrangling can stay off the to-do list, at least for this service.
What to watch
- Whether AWS publishes comparable results across more tasks, datasets and base models.
- How full fine-tuning affects general capabilities as well as task-specific scores.
- Whether teams find the serverless workflow practical at larger training scales.
Discussion spark: If full fine-tuning is easier to run, should teams use it whenever they have enough data, or does LoRA remain the better default for most custom models?
Sources and evidence
- Full Fine-Tuning (FFT) on SageMaker AI Serverless Model Customization (19 November 2026, 00:15 UTC)
- How Stretto built an AI voice agent with Amazon Nova 2 Sonic – Amazon Web Services (AWS) (19 November 2026, 00:15 UTC)
not affiliated with or endorsed by Amazon Web Services (AWS)