Amazon SageMaker AI introduced new training session APIs to support custom LLM training loops and serverless model customization without managing GPUs

Feature update Provisional 86% confidence first seen 22 October 2026, 18:00

Amazon SageMaker AI added training session APIs that let developers implement their own LLM training logic while SageMaker performs GPU-bound tasks such as distributed execution, gradient updates, checkpointing, and related operations. The service is positioned for serverless fine-tuning and reinforcement learning workflows, with support for writing custom loss/reward or advantage logic, and it uses token-based billing for compute usage.