Amazon Web Services launches SageMaker AI training session APIs

Amazon Web Services (AWS) has announced the general availability of Amazon SageMaker AI training session APIs, a serverless feature designed for custom AI model customization. The service allows developers to write their own training logic and post-training loops while AWS manages the underlying GPU infrastructure, scaling, and distributed execution.

Supported techniques include supervised fine-tuning (SFT), reinforcement learning (RL), multi-turn RL, and custom loss functions. By exposing core operations such as forward_backward() and optim_step() as API calls, the service provides fine-grained algorithmic control without the need for manual hardware provisioning. The system handles model and LoRA loading, fault recovery, and checkpoint storage.

To assist developers, AWS is also releasing the ‘SageMaker training sessions cookbook,’ which provides ready-to-run recipes for SFT and RL. The feature supports various open-weight models, including Llama, Gemma, Qwen, Nemotron, and gpt-oss. At launch, the service is available in US East (N. Virginia), US West (Oregon), and Asia Pacific (Mumbai).