Amazon Web Services (AWS) has formally launched a 'training sessions' feature in its Amazon SageMaker AI machine-learning platform, VentureBeat reported on Oct. 22.

The training section lets developers set their own model training methods, while AWS handles GPU operations.

Developers design training procedures and set success criteria, the report said. SageMaker assigns computing resources and coordinates tasks across multiple processors. It also recovers from failures and saves training progress. Companies can test their own feedback methods without building GPU infrastructure themselves.

SageMaker had already supported model fine-tuning. AWS also introduced serverless customization in December 2025. Training sessions add more granular programming control to that serverless approach, allowing developers to direct individual tasks through APIs, such as response generation and update calculation and application.

VentureBeat said the feature is useful for agent training that operates in multiple stages.

Billing is based on processed tokens rather than reserving GPU capacity. Customers are not billed while they evaluate results with their own code.

The supported models are gpt-oss-20b, Gemma-4-31B-it, Qwen3.6-27B and Qwen3.8-27B. Training data and checkpoints remain in customers' S3 storage. AWS said it does not use customer data or customized weights to train other models.

VentureBeat said services such as Thinking Machines Lab's 'Tinker' already exist in this market.