AWS Launches SageMaker AI Training Sessions for Model Customization Without GPU Management

Amazon Web Services announced the general availability of training sessions, a new capability within Amazon SageMaker AI serverless model customization, according to an AWS blog post. The feature is aimed at companies customizing open-weight foundation models using post-training techniques such as reinforcement learning.

What AWS says it does

AWS says training sessions give users control over post-training logic without requiring them to manage the underlying infrastructure. According to the company, users write and coordinate the training loop themselves, including custom loss functions or reward signals, while SageMaker AI handles GPU allocation, distributed execution, fault recovery, and checkpoint storage.

AWS states that with SageMaker AI serverless model customization, users do not need to choose instance types, manage clusters, or write distributed training configuration. The company says users can apply techniques including supervised fine-tuning (SFT), direct preference optimization (DPO), reinforcement learning with verifiable rewards (RLVR), reinforcement learning from AI feedback (RLAIF), or multi-turn reinforcement learning. AWS says users can also choose how much of a model to update using Low-Rank Adaptation (LoRA) or full fine-tuning, and can chain techniques together through what AWS calls continuous customization.

To help users get started, AWS says it has published a SageMaker AI training sessions cookbook containing examples, use cases, and model customization agent skills. According to AWS, users can describe their use case to an agent that handles data preparation, technique selection, training, evaluation, and deployment, and that this agent works within SageMaker Studio, Kiro, Claude Code, Codex, or GitHub Copilot.

Multi-turn reinforcement learning

AWS says one focus of the update is multi-turn reinforcement learning, which the company says is needed for agents that complete tasks such as booking travel, triaging support tickets, or running multi-step research, since these require several rounds of interaction before an outcome can be judged. AWS says SageMaker AI already offered multi-turn RL as a managed, serverless technique, but that training sessions were built in response to customers who wanted to define their own loss functions, reward signals, or RL stages within a single loop.

According to AWS, in a typical reinforcement learning pass, a user's training loop asks a sampler to generate several responses to a prompt using the current model, the user's reward function scores each response using a verifier, reward model, or test harness, the loop sends the scores to a trainer that calculates gradients and updates model weights, and the sampler is then refreshed with the updated weights before the next batch begins.

AWS says billing is based on the tokens SageMaker AI processes, that the billing meter stops while a user's reward function scores a batch within the user's own environment, and that there is no capacity to reserve.

Customer comments cited by AWS

AWS included a statement from Rob Gontarek, Senior Vice President of Data Science at ImagineSoftware, who said the company used the serverless training sessions to build a healthcare claim-denial triage model and that the platform's low-level API surface gave its scientists fine-grained control to pursue reinforcement learning for an accuracy-critical use case.

AWS also cited a statement from Nitin Wagh, CEO of Karini AI, who said the company fine-tuned the gpt-oss-20b model for agentic, multi-turn tool calling against its Amazon Neptune knowledge graph using the service, and that setup required no manual tuning or out-of-memory troubleshooting, with reliable checkpoint-and-resume functionality.

Source: AWS Machine Learning Blog, "Announcing new ways to customize models with SageMaker AI, without managing GPUs," published October 22, 2026.