Quick answer
In SageMaker model customization, full fine-tuning (FFT) updates all model parameters; LoRA is parameter-efficient and updates a smaller adaptation. AWS now supports FFT for more than 25 open-source models, but the exact model-and-technique matrix is not uniform.
Start with LoRA when you want a smaller, faster experiment for a bounded behavior change. Test FFT when the use case needs deeper adaptation to terminology, task structure, output formats, reasoning patterns, or domain knowledge—and when you can support the larger training and regression burden.
Neither method is automatically more accurate, cheaper, or deployable for your workload.
Decision table
| Question | LoRA starting point | Full fine-tuning starting point |
|---|---|---|
| Change depth | Narrow style or behavior adaptation | Broader model adaptation |
| Experiment speed | Prefer a lower-scope first test | Accept a larger training job |
| Data maturity | Limited but curated task data | Larger, diverse, rights-cleared corpus |
| Regression tolerance | Preserve more base behavior | Measure broad capability drift |
| Artifact strategy | Manage an adapter and compatible base | Manage a complete customized model |
| Support | Check LoRA cell for model + technique | Check FFT cell for model + technique |
AWS’s matrix separates SFT, DPO, RLVR, and RLAIF, each with LoRA and FFT columns. A model supporting SFT with FFT does not prove it supports DPO or RLVR with FFT.
Run a fair comparison
- Pin one base model, Region, dataset, and evaluation holdout.
- Choose the same customization objective and acceptance threshold.
- Use documented starting recipes; record every hyperparameter difference.
- Measure task quality, regressions, safety, training usage, artifact size, and serving behavior.
- Compare total cost per approved output, not job duration alone.
- Repeat enough runs to distinguish a stable result from sampling noise.
- Approve the simplest method meeting the task and governance requirements.
Do not use training loss as the deployment decision. Include human-reviewed cases, base-model comparisons, adversarial examples, and production-like latency and throughput tests.
Related decisions
Check the current models and Regions guide before designing the experiment. Use the setup guide for IAM and submission, and the deployment guide after evaluation.
Frequently asked questions
What is the difference between full fine-tuning and LoRA in SageMaker?
Full fine-tuning updates all model parameters, while LoRA is parameter-efficient and updates a smaller learned adaptation. The available choice depends on the exact model and customization technique in AWS’s current support matrix.
When should I test full fine-tuning first?
Test it when the workload requires deeper adaptation and you have enough high-quality data, evaluation coverage, budget, and regression controls to justify changing all parameters.
Is LoRA always cheaper or better?
No universal result follows from the method name. Compare both supported configurations on the same data, evaluation set, training budget, deployment target, and accepted-output metric.
Official sources
- AWS full fine-tuning announcement
- SageMaker supported models and customization types
- SageMaker model customization concepts
Source check: August 4, 2026. Verify the current model-by-technique matrix, training recipes, prices, quotas, licenses, and deployment compatibility before choosing a method.