Amazon Web Services Announces Training Sessions for SageMaker AI
AWS announced the SageMaker AI serverless model customization feature, which allows developers to create custom training loops without managing GPU infrastructure. Offering token-based pricing, the platform supports techniques such as RL, SFT, and DPO.
Amazon Web Services (AWS) has generally released serverless model customization and training session APIs as part of its SageMaker AI platform. This new feature allows developers to customize open-weight models using methods such as reinforcement learning (RL), supervised fine-tuning (SFT), and direct preference optimization (DPO) without the necessity of managing GPU clusters.
The new platform automates complex processes such as GPU allocation, distributed execution, gradient synchronization, and failure recovery, enabling developers to define their own training loops, custom loss functions, and reward signals. Managed through the SageMaker Python SDK with three different clients—ServiceClient, TrainingClient, and SamplingClient—the system is billed based on the amount of tokens processed rather than hourly GPU usage.
While supporting models such as gpt-oss-20b, gemma-4-31B-it, and the Qwen3 series, the platform integrates with development environments like SageMaker Studio and GitHub Copilot. Early adopters such as ImagineSoftware and Karini AI have started using the system in healthcare and multi-turn tool-calling processes. Additionally, a training cookbook featuring techniques like GRPO has been made available to developers.