Amazon announces SageMaker sessions for custom model training

Amazon announced general availability of SageMaker AI training sessions on October 22, 2026. The company says developers can write their own post-training procedure—further training an existing model for a task—without setting up or operating the graphics processing units (GPUs) that perform the heavy calculations.

SageMaker already offered managed customization methods in which users choose a model, supply data and adjust the available training settings. Amazon says the new sessions serve developers who need to change the training procedure itself, rather than fit their work into a prepackaged method.

Amazon says developers control training steps while SageMaker runs GPUs

A training loop is the repeated sequence of steps used to improve a model. According to Amazon, developers run that loop in their own Python environment and call SageMaker’s application programming interfaces (APIs) for operations that need GPUs. The model stays loaded on the service’s GPUs between calls.

Developers can define custom loss functions, which measure training errors, or reward functions, which score outputs. Their code can also connect a model to tools or simulators. This gives them control over how responses are generated, judged and used in the next model update.

Amazon illustrates the division of work with a reinforcement-learning loop, in which scores guide learning:

1. Generate responses: the developer’s code asks a sampler running on managed GPUs to produce answers from the current model.
2. Score responses: the developer’s reward function runs in their own environment, using a verifier, another model or a test harness.
3. Update the model: forward_backward() calculates gradients—the calculations that guide changes to the model’s numerical weights—and optim_step() applies the update on the GPUs.