Customers write post-training loops, and AWS handles GPU allocation, distributed execution, and fault recovery.
AWS added training sessions to Amazon SageMaker AI serverless model customization, and the feature is now generally available. Customers write and coordinate the post-training loop, including custom loss functions and reward signals, while SageMaker AI runs the underlying hardware.
Companies that get results from AI customize open-weight models instead of relying only on off-the-shelf foundation models. AWS built serverless model customization to skip instance selection, cluster management, and distributed training configuration, and the new sessions extend that work.
Teams can pick supervised fine-tuning, DPO, RLVR, RLAIF, or multi-turn reinforcement learning, and they can apply LoRA or full fine-tuning to control how much of a model changes. ImagineSoftware uses the platform for a healthcare claim-denial triage model that needs fine-grained control over forward and backward passes.
One open question is how well multi-turn reinforcement learning scales, because agent training loops grow complicated fast. AWS also ships a cookbook and model customization agent skills that run inside SageMaker Studio, Kiro, Claude Code, Codex, and GitHub Copilot, so watch which surface teams adopt first.
What matters
- AWS made training sessions generally available in SageMaker AI serverless model customization.
- Builders write the training loop while AWS handles GPU allocation, fault recovery, and checkpoints.
- Watch whether multi-turn reinforcement learning and agent skills reach production for enterprise teams.
Why it matters
Watch whether multi-turn reinforcement learning and agent skills reach production for enterprise teams.
This GenAI News article was prepared in original wording using reporting and materials published by AWS Machine Learning Blog. Source reference: https://aws.amazon.com/blogs/machine-learning/announcing-new-ways-to-customize-models-with-sagemaker-ai-without-managing-gpus/.
Drafted by the GenAI News review pipeline.