SageMaker AI training sessions reach general availability for custom model training

SageMaker AI serverless model customization already removes the need to pick instance types, manage clusters or write distributed training configuration. The new training sessions let developers write and coordinate the training loop themselves, including custom loss functions and reward signals. Supported techniques include SFT, DPO, RLVR, RLAIF and multi-turn RL, with either LoRA or full fine-tuning.

In each pass, the developer's code asks a sampler to generate several responses per prompt and scores them with a reward function running in the developer's own environment. The scores go to a trainer that updates the model weights through forward_backward() and optim_step(), and the session then refreshes the sampler with the new weights. Billing covers the tokens SageMaker AI processes, and the meter stops while the reward function scores a batch.

Each session runs under an IAM execution role the customer creates and can use the customer's own AWS KMS key and VPC configuration. Data and checkpoints stay in the customer's Amazon S3 buckets in the chosen Region, and AWS says it does not use customer data or customized weights to train other models. Developers can install the SageMaker Python SDK on Linux or macOS with Python 3.10 to 3.12, or work in SageMaker Studio with coding agents such as Kiro, Claude Code, Codex or Copilot.

Teams building agents can run multi-turn reinforcement learning with their own reward functions and loss definitions without managing GPU clusters. Billing is based on processed tokens, and the meter stops while the reward function scores a batch in your environment. Check that your base model is one of the four listed models before planning a training run.