Amazon Web Services has made training sessions in Amazon SageMaker AI generally available, letting developers write and coordinate their own post-training loops while AWS manages the GPUs underneath. The company announced the general availability on October 22, expanding its serverless model customization offering.
Training sessions give developers control over the post-training logic without managing infrastructure. Teams write the training loop — including custom loss functions or reward signals — and SageMaker AI handles GPU allocation, distributed execution, fault recovery and checkpoint storage. The target is the hardest part of model customization: techniques like reinforcement learning and multi-turn RL that need environments, iteration and careful orchestration.
What training sessions are
SageMaker AI serverless model customization already offered managed techniques — supervised fine-tuning, DPO and other reinforcement learning methods — where teams pick a technique and a model and let SageMaker run the training. Training sessions go further: developers get to write the training loop itself, whether that means reproducing SFT or DPO on their own terms or implementing a custom loss, a new RL reward signal, a data curriculum or an algorithm straight from the latest research.
Operations like forward_backward(), optim_step() and sample() become API calls against a live, stateful session, so customers can compose whatever training procedure their technique requires. Teams can train from labeled examples, preferences or rewards, use LoRA or full fine-tuning, and chain techniques. Launch support covers gpt-oss-20b, gemma-4-31B-it, Qwen3.6-27B and Qwen3.8-27B.
How they work
Instead of a single fire-and-forget job, developers get API endpoints to manage the lifecycle of a training run: start, check status, stop. In AWS’s GRPO example, a customer loop gathers model responses, scores them locally, returns the scores to a trainer and refreshes the sampler after updates.
AWS is also shipping a SageMaker training sessions cookbook with ready-to-run SFT and RL recipes as blueprints, aimed at developers across experience levels. Access runs through Python SDK clients from a local machine or from SageMaker Studio, and the company suggests connecting a coding agent to drive the loop.
Billing and availability
Billing is per token processed, and local computation — like reward scoring — is not billed as GPU time. That matters for RL workflows, where much of the loop’s cost sits in scoring rather than gradient updates. The per-token model also means teams pay for what the model processes, not for reserved capacity sitting idle between experiments.
Training session APIs are generally available in US East (N. Virginia), US West (Oregon) and Asia Pacific (Mumbai). AWS’s customer example is Karini AI, presented as a testimonial rather than an independent performance test — the real measure will be how teams use the cookbook recipes in production.
Why It Matters
Post-training is where models become products: the step that turns a base model into something that follows a company’s workflow, respects its policies and knows its domain. It has also been the step that demanded the most infrastructure babysitting. By abstracting the GPUs while handing developers the training loop itself, AWS is aiming at the teams building specialized agents — the ones for whom off-the-shelf fine-tuning is not enough and research-grade flexibility is the requirement.
The launch also sharpens the cloud AI contest. Every hyperscaler is trying to own the model-customization layer: Google with Vertex AI tuning, Microsoft with Azure AI Foundry, AWS now with training sessions. The winner gets the developers who are turning open-weight models into enterprise systems.
Background
Amazon SageMaker AI is AWS’s platform for building, training and deploying machine learning models, and serverless model customization has been its answer to the operational cost of fine-tuning. Training sessions extend that bet: the infrastructure disappears into the API, and what remains is the training procedure itself — written by the developer, run by the cloud.
Source: AWS Machine Learning Blog (official announcement) ; VentureBeat
0 Comments