SageMaker Serverless Model Customization Cuts Setup
Amazon SageMaker AI now offers serverless model customization, replacing manual cluster provisioning with token-based pricing and multi-technique post-training.
Amazon SageMaker AI serverless model customization allows developers to launch a ready-to-use AI development environment in SageMaker Studio in under 20 seconds. Under this workflow, users do not choose instance types, manage clusters, or write distributed training configurations.1
Serverless execution in Amazon SageMaker AI

During training sessions, SageMaker AI automates backend operations by handling GPU allocation, distributed execution, fault recovery, and checkpoint storage. Teams evaluating LLM API pricing or architectural tradeoffs in agent reinforcement learning avoid provisioning dedicated clusters for iterative runs.13
Supported models and post-training methods
The serverless service supports four core foundation models: gpt-oss-20b, gemma-4-31B-it, Qwen3.6-27B, and Qwen3.8-27B. Practitioners can customize these models using multiple post-training techniques, including Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), Reinforcement Learning with Verifiable Rewards (RLVR), Reinforcement Learning from AI Feedback (RLAIF), and multi-turn Reinforcement Learning (RL).1
Tuning depth and training logic control
Users retain granular control over how much of the model is updated, choosing between parameter-efficient Low-Rank Adaptation (LoRA) and full fine-tuning. Teams can also chain techniques together through continuous customization. Training sessions allow engineers to write and coordinate the training loop directly, retaining full authority over post-training logic such as custom loss functions and reward signals without operating the underlying compute nodes.1
Token-based pricing and S3 data retention
Billing operates under a pay-as-you-go pricing model with no upfront commitments or minimum fees, charging specifically for the tokens processed by SageMaker AI. The meter automatically stops while a reward function scores a batch in the user's environment, eliminating the need to reserve GPU capacity ahead of time.12
All training data remains strictly inside the user's own Amazon Simple Storage Service (Amazon S3) buckets within their selected AWS Region. Training checkpoints generated during execution land directly in those same designated S3 buckets.1