Quick answer
AWS announced SageMaker serverless full fine-tuning for more than 25 open-source models from the gpt-oss, Gemma, Llama, Nemotron, and Qwen families. The launch Regions were:
- US East (N. Virginia);
- US West (Oregon);
- Asia Pacific (Tokyo);
- Europe (Ireland).
Do not turn that family list into a blanket support claim. The current developer guide publishes an exact model ID matrix with separate cells for SFT, DPO, RLVR, and RLAIF using LoRA or full fine-tuning. Some model-and-technique combinations are blank.
Resolve the exact combination
Record this tuple before creating data or infrastructure:
AWS Region
model provider and display name
SageMaker model ID
customization technique
training type: LoRA or FFT
model license and acceptable-use terms
dataset location
evaluation location
deployment targetA family name is not a model ID. A supported model is not necessarily supported for every technique. Training support also does not prove that the chosen artifact can be deployed through either target in the same Region.
Region review
Keep the training job, input data, outputs, logs, evaluations, model package group, and deployment plan in one reviewed Region map. The SageMaker customization overview notes that features powered by Amazon Bedrock Evaluations may securely transmit data across AWS Regions within the same geography for processing. Review that behavior against your data-location requirements rather than assuming every evaluation is single-Region.
Check service availability, quotas, IAM, KMS keys, S3 bucket policies, VPC controls, and deployment support for the exact Region. Re-run this review when a model or technique changes.
Avoid a stale support table
The official support matrix changed recently enough that copying every row into an evergreen guide would create unnecessary drift. Use AWS’s live table as the source of truth and preserve the date in your decision record.
After resolving support, use the FFT versus LoRA guide and full fine-tuning setup.
Frequently asked questions
Which model families did AWS announce for serverless full fine-tuning?
AWS announced more than 25 open-source models across gpt-oss, Gemma, Llama, Nemotron, and Qwen families. The developer guide contains the current model IDs and technique matrix.
Which Regions launched SageMaker serverless full fine-tuning?
AWS listed US East (N. Virginia), US West (Oregon), Asia Pacific (Tokyo), and Europe (Ireland) in the August 3, 2026 announcement.
Does every listed model support every customization technique?
No. AWS publishes separate support cells for SFT, DPO, RLVR, and RLAIF with LoRA or full fine-tuning. Verify the exact model ID and cell before designing a job.
Official sources
- AWS full fine-tuning announcement
- SageMaker supported models and customization types
- SageMaker model customization overview
Source check: August 4, 2026. Treat the AWS support matrix, Regional availability, quotas, deployment compatibility, and model terms as live facts that must be rechecked.