SageMaker AI adds serverless training sessions for custom post-training

SageMaker AI serverless model customization now includes general availability of training sessions. Users write and coordinate the training loop with custom loss functions or reward signals. SageMaker AI manages GPU allocation, distributed execution, fault recovery, and checkpoint storage without requiring users to choose instance types or manage clusters.

Supported techniques include SFT, DPO, RLVR, RLAIF, and multi-turn RL. Users can apply LoRA or full fine-tuning and chain techniques. The service supports gpt-oss-20b, gemma-4-31B-it, Qwen3.6-27B, and Qwen3.8-27B. Access is available today through the SageMaker Python SDK or SageMaker Studio.

Data stays in customer S3 buckets in the chosen AWS Region. Sessions use an IAM execution role, optional KMS keys, and VPC configuration. AWS states it does not use customer data or customized weights to train other models. Users pay only for tokens processed; the meter pauses during local reward scoring.

A cookbook with examples and agent skills is provided for use in SageMaker Studio, Kiro, Claude Code, Codex, or GitHub Copilot. The agent can handle data prep, technique selection, training, evaluation, and deployment based on a described use case.