AWS SageMaker AI launches training session APIs for serverless model customization

Amazon SageMaker AI on October 22, 2026 made training session APIs generally available for its serverless model customization offering. Developers can write custom training loops using calls like forward_backward(), optim_step() and sample() without provisioning GPU infrastructure. The APIs are available in three regions and billed per token processed.

• APIs available in US East (N. Virginia), US West (Oregon) and Asia Pacific (Mumbai)
• SageMaker handles GPU allocation, LoRA weight loading, distributed execution and checkpoints
• Training sessions are billed per token processed, not per GPU hour
• A cookbook with ready-to-run SFT and RL recipes was released