AWS makes SageMaker AI Training Sessions generally available for custom model development

Amazon Web Services made SageMaker AI Training Sessions generally available on October 22, 2026, giving development teams a serverless way to write their own AI model training routines without managing the underlying GPU hardware.

The new capability sits between AWS's preset fine-tuning options and its full infrastructure tools. Developers write the training logic — including how the model is scored and how it updates its own parameters — while SageMaker handles GPU allocation, distributed execution across processors, fault recovery, and checkpoint storage (saving training progress at intervals). The service supports five post-training methods, including supervised fine-tuning and reinforcement learning, and launches with four named open-weight base models: gpt-oss-20b, gemma-4-31B-it, Qwen3.6-27B, and Qwen3.8-27B. Billing is token-based, covering input processing, response generation, and training steps separately. Teams also need to budget for any storage, evaluation services, and production inference used alongside the training sessions.

For IT directors and data teams that have been holding off on custom model development because of GPU infrastructure costs and complexity, this is a meaningful change in the practical barrier to entry. Organizations in regulated or specialized fields — healthcare, legal, financial services, manufacturing — that want a model trained on their own terminology and workflows now have a more accessible path to do that inside an existing AWS environment. Teams that are evaluating the option may want to review which base models are currently supported and confirm that their intended training approach is covered before committing to a project timeline.