AWS launches Training Sessions for SageMaker AI

Amazon Web Services launched "Training Sessions," a new feature of Amazon SageMaker AI, on October 22 local time. The feature lets developers define training procedures and evaluation criteria in code while AWS handles GPU allocation, distributed training, fault recovery and checkpointing, according to aitimes.com.

Developers can customize how an AI generates answers, evaluates results and updates the model, including custom reward functions and multi-step training. AWS gave an example using Group Relative Policy Optimization (GRPO), where an evaluation function scores multiple answers and the model is updated based on relative scores. Supported customization ranges from LoRA to full fine-tuning.

Initial users cited by AWS are ImagineSoftware, which built a model to classify medical claim denials, and Karini AI, which customized GPT-OSS-20B for an agent interacting with the Amazon Neptune knowledge graph. AWS did not disclose specific figures on accuracy gains or total training cost per customer.

Pricing is based on processed tokens rather than reserved GPU time, and billing pauses while developer-written evaluation code reviews results. Supported models include GPT-OSS-20B, Gemma-4-31B-it, Qwen3.6-27B and Qwen3.8-27B. The feature is accessible via SageMaker Studio or the Python SDK.

Customer training data and model checkpoints are stored in Amazon S3 in the customer-chosen region, and customer data is not used to train other models. Previously, SageMaker AI supported model fine-tuning and reinforcement learning with custom training code, but developers had to configure GPU instance types, training environments and distributed training settings themselves. In December 2025, AWS introduced a serverless model customization feature that reduced infrastructure management burden but limited fine-grained control over the training process.