Announcing new ways to customize models with SageMaker AI, without managing GPUs
Amazon Web Services Pooja Karadgi ● Covered by 2 sources
AWS added serverless training sessions in SageMaker AI so teams can customize models without managing GPUs. It gives users more control over post-training, while AWS handles the infrastructure and checkpointing.
Based on reporting by Amazon Web Services, Pooja Karadgi — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
AWS is pitching a simpler way to tune AI models without taking on the usual GPU overhead. The company says SageMaker AI serverless model customization now includes general availability for training sessions, letting teams write the training loop and keep control over the post-training logic while AWS runs the infrastructure behind it.
That matters because customization is where a lot of the real value shows up. AWS points to models trained on company data, product catalogs, or clinical vocabulary as being better suited to their jobs and cheaper to run. But the path there usually means choosing a technique, wiring up GPU infrastructure, and, for agents, building out multi-turn reinforcement learning environments. Training sessions are meant to remove that plumbing.
The promise is broad. Customers can use methods like SFT, DPO, RLVR, RLAIF, or multi-turn RL, and they can choose between LoRA and full fine-tuning. AWS also says those techniques can be chained through continuous customization. On top of that, the company is pushing a cookbook and model-customization agent skills that can handle data prep, technique selection, training, evaluation, and deployment.
AWS is clearly aiming at more hands-on teams, not just people who want a button. In one example, ImagineSoftware used SageMaker AI Serverless Training Sessions to build a healthcare claim-denial triage model. The company said the low-level API, along with control over forward and backward passes, sampling, and the optimizer step, gave its scientists the control they wanted for a reinforcement-learning use case.
Another focus is agent training, especially multi-turn workflows. AWS says these jobs often need several rounds of interaction before anyone can tell whether the model actually succeeded. Training sessions keep the loop flexible: your code handles the reward function and environment, while SageMaker AI does the GPU-heavy parts such as sampling, gradient calculation, weight updates, fault recovery, and checkpoint storage. AWS also says you pay for the tokens it processes, with the meter stopping while your reward function scores a batch.
The company is also leaning hard on data control. Each session uses an IAM role you create, and you can add your own KMS key and VPC setup. Data and checkpoints stay in your S3 buckets in the AWS Region you choose, and AWS says it won’t use your data or customized weights to train another model. Support is listed today for gpt-oss-20b, gemma-4-31B-it, Qwen3.6-27B, and Qwen3.8-27B.
My take — AI-written commentary, not fact-checked reporting
This is AWS doing what AWS does best: turning painful infrastructure into a managed service and calling it progress, because for most teams it is. The real story is not the model names, it’s the shift toward controlled post-training without a GPU babysitting job. Open-weight models keep winning when the platform makes them practical instead of theatrical.
Read more about this at: Amazon Web Services