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

Amazon made training sessions generally available in SageMaker AI serverless model customization, letting teams write their own post-training loop while AWS handles GPU allocation, distributed execution, fault recovery and checkpoint storage.

• The company says companies getting real results from AI customize open-weight models with techniques such as reinforcement learning rather than using foundation models off the shelf, but that customization requires picking a technique and managing GPU infrastructure.
• Training sessions let customers define custom loss functions, reward signals and RL stages in one loop, and support SFT, DPO, RLVR, RLAIF, multi-turn RL, LoRA or full fine-tuning, chained through continuous customization.
• A cookbook with examples and model customization agent skills works in SageMaker Studio, Kiro, Claude Code, Codex or GitHub Copilot.
• ImagineSoftware's Rob Gontarek said the low-level API surface and direct control over forward and backward passes, sampling and the optimizer step gave scientists the control needed for a healthcare claim-denial triage model.
• Karini AI's Nitin Wagh said the platform handled infrastructure while fine-tuning gpt-oss-20b for multi-turn tool calling against its Amazon Neptune knowledge graph, with reliable checkpoint-and-resume.
• Customers pay for tokens processed, the meter stops while a reward function scores a batch, and there is no capacity to reserve; each session uses an IAM role, with optional KMS key and VPC configuration, and data and checkpoints stay in the customer's S3 buckets.