Skip to content
NewsProduct launch

AWS takes SageMaker training sessions to GA, letting teams write their own RL loops without touching a GPU

· by Pondero Newsdesk

The short version

AWS announced general availability of training sessions for Amazon SageMaker AI, a serverless layer that lets teams write custom post-training and reinforcement-learning loops while AWS manages the GPU infrastructure underneath.

AWS takes SageMaker training sessions to GA, letting teams write their own RL loops without touching a GPU

Teams building agents that need multi-turn reinforcement learning have had to either accept a vendor's fixed training recipe or manage their own GPU clusters to customize one. AWS closed that gap on October 22, 2026, announcing general availability of training sessions for Amazon SageMaker AI, a serverless layer where customers write the training loop itself and AWS handles everything underneath it, per AWS's announcement post.

What shipped

Training sessions give developers direct control over post-training logic, including custom loss functions and reward signals, while SageMaker AI handles GPU allocation, distributed execution, fault recovery, and checkpoint storage, according to the announcement. The feature supports standard techniques (SFT, DPO, RLVR, RLAIF) plus multi-turn reinforcement learning for agents that need several rounds of interaction, such as an agent booking travel or triaging support tickets, before a result can be scored. Teams can update a model fully or with LoRA, and chain techniques together through continuous customization. At GA, training sessions support four open-weight models: gpt-oss-20b, gemma-4-31B-it, Qwen3.6-27B, and Qwen3.8-27B, per the companion technical post. Billing is metered by tokens processed, and the meter pauses while a customer's own reward function scores a batch outside SageMaker's infrastructure, so there is no idle-GPU charge during that step and no capacity to reserve upfront.

Builders keep the loop, AWS keeps the clusters

The pitch here is narrower than a typical fine-tuning product launch: AWS already offered managed multi-turn RL as a fixed, serverless technique, but said customers wanted to define their own loss functions and reward shaping inside that same loop rather than accept the built-in recipe. Training sessions answer that specific complaint without asking the customer to provision instance types, build distributed training configuration, or run separate inference and training clusters by hand. ImagineSoftware, a healthcare billing vendor, said in the AWS post that the feature's low-level API access to the forward and backward pass, sampling, and the optimizer step let its data scientists pursue RL on a claim-denial triage model, a use case AWS describes as "accuracy-critical." Data, checkpoints, and model weights stay in the customer's own Amazon S3 buckets under an IAM role and KMS key the customer controls, and AWS states it does not use that data or the resulting weights to train other models, a claim that is AWS's own and not independently verified here.

What to watch next

Watch whether AWS expands the four-model GA list (gpt-oss-20b, Gemma 4 31B, and the two Qwen 3.6/3.8 27B variants) to larger or additional model families, since the current roster caps out around 31B parameters. Also worth tracking is whether independent engineering teams publish cost or reliability comparisons against self-managed GPU training pipelines for the same multi-turn RL workloads, since AWS has not published third-party benchmarks of its own.

Sources