Announcing new ways to customize models with SageMaker AI, without managing GPUs

AWS announces general availability of SageMaker AI training sessions, enabling serverless model customization with custom post-training loops without managing GPU infrastructure.

• Write custom post-training logic including SFT, DPO, RLVR, RLAIF, and multi-turn reinforcement learning without infrastructure management
• Training sessions handle GPU allocation, distributed execution, fault recovery, and checkpoint storage automatically
• Supports techniques like Low-Rank Adaptation (LoRA), full fine-tuning, and chaining multiple techniques together
• AI agent assistance available in SageMaker Studio, Kiro, Claude Code, Codex, and GitHub Copilot for data prep and model evaluation
• Pay only for tokens processed; meter stops during reward function scoring with no capacity reservation required
• Data and customized models remain in customer's S3 buckets with IAM, KMS, and VPC security controls
• Supports gpt-oss-20b, gemma-4-31B-it, Qwen3.6-27B, and Qwen3.8-27B models

SageMaker AI training sessions simplify model customization by eliminating infrastructure management while providing fine-grained control over training loops and reinforcement learning workflows.