Amazon SageMaker AI has made serverless training sessions generally available, giving developers control over post-training logic while AWS manages the GPU infrastructure. Teams write the loop and scoring rules rather than selecting instances or configuring distributed clusters.
Supported approaches include supervised fine-tuning, preference optimization, reinforcement learning and Low-Rank Adaptation or full-model updates. For a multi-turn agent, a developer’s environment can run tools and calculate rewards while SageMaker’s sampler generates responses and its trainer applies gradient updates. Checkpoints and fault recovery are handled by the service.
Billing is based on tokens processed by SageMaker, and the meter pauses while an external reward function scores a batch. Data and checkpoints stay in the customer’s S3 buckets in the selected region. Sessions use a customer-created IAM role and can add a KMS encryption key and virtual private cloud configuration.
The initial supported models are gpt-oss-20b, gemma-4-31B-it, Qwen3.6-27B and Qwen3.8-27B. AWS also provides a cookbook and agent skills for setup, evaluation and deployment. The managed infrastructure reduces operational work, but customers still own training logic, reward quality, safety evaluation and the decision to deploy a customized model.