The service lets customers write and coordinate their own post-training loop while AWS manages GPU allocation, distributed execution, fault recovery, and checkpoint storage.
AWS says users can supply custom loss functions and reward signals, including for multi-turn reinforcement learning. The announced workflow supports techniques including supervised fine-tuning, DPO, RLVR, RLAIF, and multi-turn RL, with either LoRA or full fine-tuning. AWS also describes continuous customization, which can chain techniques together.
The company says its cookbook and model-customization agent skills can assist with data preparation, technique selection, training, evaluation, and deployment. It lists SageMaker Studio, Kiro, Claude Code, Codex, and GitHub Copilot as working environments. The announcement does not establish pricing, regional availability, compute limits, supported-model coverage, or independently measured quality or cost results. Source: https://aws.amazon.com/blogs/machine-learning/announcing-new-ways-to-customize-models-with-sagemaker-ai-without-managing-gpus