AWS adds custom AI training loops to SageMaker without customer-managed GPUs

AWS is taking GPU management out of custom model training—not the work of deciding how a model should learn. On October 22, 2026, it made SageMaker AI serverless model customization training sessions generally available. Developers write their own training loops and reward signals, while AWS handles the computing infrastructure behind them.

The training recipe stays with the developer

The release extends an existing service rather than introducing serverless customization from scratch. SageMaker AI already offered managed multi-turn reinforcement learning: training in which an agent takes several actions before its result is scored. AWS says customers wanted more control over the rules used to calculate learning updates, the rewards assigned to responses, and the stages within a training loop.

Training sessions give developers that lower-level control. Their code runs the training loop and reward function in their own environment, connecting an agent to tools or simulators. AWS allocates GPUs, coordinates work across machines, recovers from failures and stores checkpoints—saved training states. Customers do not select instance types, manage clusters or write distributed-training configuration.

The supported approaches include learning from labeled examples, preferences or rewards, including rewards for verifiably correct results. Developers can use Low-Rank Adaptation, or LoRA, to adjust a limited set of parameters, or choose full fine-tuning to update the whole model. They can also chain techniques together rather than restrict an experiment to one method.

Responses go out; scores come back

AWS illustrates the division of work with Group Relative Policy Optimization, or GRPO. This method generates several responses to the same prompt and uses their relative scores to guide learning. The customer’s loop requests responses from a sampler, then scores them locally using a verifier, reward model or test harness. Those scores return to a trainer, which calculates and applies model updates. The loop then refreshes the sampler with the updated model and repeats.

A customer example puts that control in context. In AWS’s announcement, Karini AI CEO Nitin Wagh said his team fine-tuned gpt-oss-20b for multi-turn tool calling against its Amazon Neptune knowledge graph. He described reliable checkpoint-and-resume behavior and said the platform avoided manual tuning and out-of-memory problems. That is a customer testimonial published by AWS, not an independent performance test.

Four models, with customer-owned storage

At launch, training sessions support gpt-oss-20b, gemma-4-31B-it, Qwen3.6-27B and Qwen3.8-27B. Data and checkpoints stay in the customer’s own Amazon S3 buckets in a chosen AWS Region. Each session uses a customer-created execution role for permissions, with options to add a customer-managed encryption key and virtual private cloud configuration. AWS says it does not use customer data or customized weights to train another model.

AWS offers two starting routes, backed by a training sessions cookbook with examples and coding-agent skills:

• In a local development environment, install the SageMaker Python SDK on Linux or macOS with Python 3.10–3.12, then clone the cookbook.
• In SageMaker Studio, open a notebook and connect a coding agent such as Kiro, Claude Code, Codex or GitHub Copilot, then download the skills and cookbook.