Announcing new ways to customize models with SageMaker AI, without managing GPUs
aws ·
AWS has announced the general availability of SageMaker AI serverless model customization training sessions, allowing developers to customize models without managing GPU infrastructure. This new feature enables users to write and coordinate their own post-training loops, including custom loss functions and reward signals, while AWS handles distributed execution and fault recovery. Customers can utilize various techniques such as SFT, DPO, RLVR, and multi-turn reinforcement learning with full control over the training process. The service supports specific open-weight models including gpt-oss-20b, gemma-4-31B-it, and Qwen variants. Users retain ownership of their data and models, which remain in their own S3 buckets and are not used to train other models. Implementation is facilitated through the SageMaker Python SDK or directly within SageMaker Studio using integrated coding agents. Early adopters from healthcare and AI sectors have praised the platform for its fine-grained control and operational simplicity.
aws