What happened
AWS says SageMaker AI Serverless Model Customization now supports full fine-tuning, in addition to parameter-efficient methods such as Low-Rank Adaptation (LoRA). LoRA freezes the pretrained model weights and trains smaller low-rank matrices; FFT updates the model’s parameters directly. AWS presents FFT as an option for complex tasks with large, good-quality datasets, while noting that LoRA can remain effective and less resource-intensive. The serverless service handles compute allocation and checkpoint management: users specify a base model, datasets and hyperparameters rather than provisioning GPUs or setting up distributed training themselves.
For its comparison, AWS used supervised fine-tuning on Qwen3-4B with the Magicoder-Evol-Instruct-110K dataset, a collection of 110,000 coding instruction-response pairs. The split was 99,000 training examples, 5,500 for validation and 5,500 held out for testing. Both methods received the same prompts, targets and evaluation code. AWS says it also ran experiments across other supported base models, including DeepSeek-R1-1.5B, Nemotron-30B and Gemma-4-26B; each experiment, from job submission through evaluation, took a few hours, and fine-tuned models were registered as versioned Model Packages.
The example SageMaker Python SDK configuration sets TrainingType.FULL, a learning rate of 2e-5, global batch size 8, two epochs and a warmup-step ratio of 0.1. AWS says switching to LoRA requires changing the training type while leaving the rest of the example intact. Its evaluation combines LLM-as-Judge, PRIMECODE and MMLU: respectively, semantic correctness judged by Amazon Nova Pro, code similarity measures such as BLEU and ROUGE, and a test of general knowledge retention across 57 subjects.
AWS reports that FFT’s LLM-as-Judge semantic-correctness score reached 90.0% at three epochs in its code-generation experiments, but extending training beyond three epochs degraded every variant it tested. Its results also show trade-offs rather than an across-the-board win: LoRA scored 81.5% versus FFT’s 79.5% on one LLM-as-Judge comparison, and both methods underperformed the base model on MATH exact match. AWS identifies learning rate and dataset scale as important variables: it describes 2e-5 with more than 100,000 examples as a safer zone in its experiments, while learning rates of 5e-5 or higher sharply harmed MATH and MMLU scores. Its best general-knowledge retention configuration used 5% warmup, batch size 8 and a 2e-5 learning rate.
What the archive adds
New Runtime covered SageMaker AI model-customization training sessions that require no GPU management in an earlier note dated 2026-10-22. This announcement adds a distinct training method—full fine-tuning—and AWS’s comparison with LoRA, rather than merely restating the managed-training capability. A separate note about cloud fine-tuning costs and failures is related by topic, but does not establish results for this service or experiment.
Trend signal
This does not yet fit a pattern New Runtime tracks. The announcement is a managed customization feature and a vendor-run comparison of training methods; it does not establish a broader change in agent architecture, interoperability, or model economics. Its most useful signal for builders is narrower: managed training can make full-parameter experiments easier to run, while the quality-versus-forgetting trade-off still depends on dataset scale and careful evaluation.
Why it matters
For teams deciding whether to adapt a model rather than rely on prompting or a larger general-purpose model, FFT becomes a more accessible experiment on AWS. It may be worth testing when the task is complex, the training corpus is substantial and LoRA has reached an accuracy ceiling. The operational convenience does not remove the main risk: updating every weight can degrade capabilities outside the target task. AWS’s use of a held-out test set and MMLU as a forgetting check offers a useful evaluation pattern, but teams should choose checks that reflect their own production requirements. A higher score on a coding benchmark alone is not enough to justify deployment if general reasoning or other critical behaviors regress.
Limits and uncertainty
AWS’s reported results come from its own experiments on a coding dataset and selected models; they do not establish that FFT will outperform LoRA on other tasks, datasets or evaluation criteria. LLM-as-Judge depends on an automated evaluator, while text-similarity scores reward overlap with reference solutions and may not capture all dimensions of code quality. The post’s hyperparameter guidance is explicitly based on those experiments, not a universal recipe. Teams still need to validate training cost, deployment behavior and capability retention on their own data and workloads.
Reader takeaway
Treat FFT as another option to test, not a default replacement for LoRA. If you have a large, high-quality dataset and need deeper adaptation, compare both methods on a held-out set, track general capabilities as well as task performance, and be cautious with learning rates and extra epochs. The serverless workflow reduces infrastructure setup; it does not make model quality or regression risks automatic to manage.