AWS launches SageMaker Training Sessions for custom AI model training

Amazon Web Services wants developers to stop babysitting GPUs. On October 22, 2026, the company announced general availability of Amazon SageMaker AI Training Sessions, a serverless feature for customizing AI models.

The pitch is simple. Developers write the training logic, and AWS manages the hardware, the scaling and the distributed plumbing underneath.

What Training Sessions actually does

In technical terms, Training Sessions lets developers run their own post-training loops. Supported approaches include supervised fine-tuning (SFT), reinforcement learning (RL), custom loss functions and multi-turn RL.

For the uninitiated, supervised fine-tuning means showing a model examples of correct answers. Reinforcement learning means rewarding it when it gets things right, so it learns which behaviors pay off. Multi-turn RL extends that reward process across longer, back-and-forth interactions, which is closer to how agents operate in practice.

Meanwhile, AWS handles GPU management, distributed execution and infrastructure scaling. The developer keeps algorithmic control without inheriting the operations headache.

Access runs through Python SDK clients, either from a local machine or from SageMaker Studio. Developers can call session operations directly, including forward_backward and optim_step.

Those two names describe the core rhythm of model training. One pass computes how wrong the model was and which direction to adjust. The other actually applies the adjustment. Exposing them directly gives developers fine-grained control over each step of the loop.

AWS is also shipping a companion “SageMaker training sessions cookbook” with ready-made recipes for SFT and RL, aimed at developers across experience levels.

Models, regions and the bill

The feature supports a range of popular open-weight models. The list includes Llama, Gemma, Qwen, Nemotron and gpt-oss.

At launch, Training Sessions is available in US East (N. Virginia), US West (Oregon) and Asia Pacific (Mumbai).

Security features include automatic encryption at rest in S3, with server-side encryption that can use customer-managed KMS keys. Access is governed through IAM managed policies.

Pricing is usage-based and tied to the volume of tokens processed, so customers pay roughly in proportion to how much training work they actually run.

How it fits into AWS’s serverless push

Training Sessions is not arriving in isolation. It builds on earlier serverless customization features inside SageMaker.

Serverless full fine-tuning arrived in August 2026. Support for custom sequence lengths followed on October 21, 2026, just one day before Training Sessions went generally available.

Full fine-tuning covered teams that wanted a more packaged experience. Training Sessions targets the more hands-on crowd, the ones who want to write custom loss functions and control every optimizer step without owning a cluster.