AWS Opens SageMaker Training Sessions So Teams Can Write Custom Model Training Loops

Amazon Web Services has made training sessions generally available in SageMaker AI serverless model customization. An AWS Machine Learning Blog post dated 22 October 2026 says developers can write and coordinate their own training loop, including custom loss functions or reward signals, without choosing instance types or managing clusters.[1]

The feature targets teams that adapt open-weight models with post-training methods such as reinforcement learning. AWS argues that a model trained on a firm’s own context is more accurate for its task and costs less to run. The post offers no benchmark for that claim.[1]

SageMaker AI already offered serverless model customization with managed techniques. Customers wanted more control, so AWS added training sessions. In a session, the customer’s code runs the training loop and the reward function in its own environment. The session performs the GPU-heavy work: a sampler generates responses and a trainer updates the model weights.[1]

AWS says the service handles GPU allocation, distributed execution, fault recovery and checkpoint storage. Supported techniques include SFT, DPO, RLVR, RLAIF and multi-turn reinforcement learning. Teams can update part of a model with Low-Rank Adaptation (LoRA) or run full fine-tuning, and can chain techniques.[1]

The documentation calls this operation-level control. A developer opens a session and calls operations such as forward_backward(), optim_step() and sample(). The model stays loaded on GPUs between calls, so AWS says there are no cold starts.[2]

The launch post names four supported models: gpt-oss-20b, gemma-4-31B-it, Qwen3.6-27B and Qwen3.8-27B. The documentation adds the regions. gpt-oss-20b runs in us-east-1, us-west-2 and ap-south-1. Gemma and Qwen3.6 run in us-west-2 and ap-south-1, while Qwen3.8 is listed only in us-west-2.[1][2]

AWS says customers pay for the tokens SageMaker AI processes and that the meter stops while their own reward function scores a batch. The documentation splits charges into prefill, sample and train and refers readers to the pricing page for rates. The default quota is two concurrent fine-tuning sessions, which can be raised.[1][2]

Each session uses an IAM role the customer creates, and customers can add their own AWS KMS key and virtual private cloud settings. AWS says data and checkpoints stay in the customer’s own Amazon S3 buckets in the chosen Region. It also says it does not use customer data or customised weights to train another model.[1]

Two customers describe early use: ImagineSoftware, on a healthcare claim-denial triage model, and Karini AI, which fine-tuned gpt-oss-20b for tool calling against an Amazon Neptune graph. Both are testimonials published by AWS, with no accuracy or cost figures.[1]