Amazon Web Services wants to take the hardware chores out of custom model training without dictating the recipe. On October 22 the company made training sessions generally available in Amazon SageMaker AI, extending the serverless model customization service it launched in December 2025. The feature lets developers write their own post-training logic, from standard fine-tuning to experimental reinforcement learning, while AWS provisions and operates the GPUs underneath.
Until now, SageMaker's serverless option worked from a menu. Customers picked a model and a technique such as supervised fine-tuning, direct preference optimization or one of several reinforcement learning methods, and AWS ran the job. Training sessions open the loop itself. A developer opens a stateful session and issues individual steps as API calls, using operations named forward_backward(), optim_step() and sample(). The model stays loaded in GPU memory between calls, and AWS takes care of allocating hardware, loading model and LoRA weights, spreading work across chips, recovering from failures and storing checkpoints.
That split matters most for agents. AWS's launch post walks through Group Relative Policy Optimization, in which the model drafts several answers to a prompt, the customer's own code scores them, and the scores drive the next update. Reward code and connections to tools or simulators run in the customer's environment, so an agent can be trained on multi-step tasks against real systems. Billing follows the tokens SageMaker processes, and AWS says the meter pauses while a batch is being scored, with no capacity to reserve.
At launch the service supports gpt-oss-20b, gemma-4-31B-it, Qwen3.6-27B and Qwen3.8-27B across three AWS regions, Asia Pacific (Mumbai), US West (Oregon) and US East (N. Virginia). VentureBeat, reviewing AWS's documentation, found that not every model is offered in every region and that the default quota allows two concurrent sessions. AWS is also publishing a cookbook of ready-made SFT and RL recipes and agent skills meant to let coding assistants, including GitHub Copilot, Codex, Claude Code and Kiro, drive data preparation, training and evaluation.
Two early customers are named in the announcement. ImagineSoftware used the API to build a reinforcement-learning model that triages denied healthcare claims, and Karini AI tuned gpt-oss-20b for multi-turn tool calling against an Amazon Neptune knowledge graph. Both praise the control and the reduced setup work, but neither quote includes accuracy, cost or speed figures. Sessions run under a customer-created IAM role, can use customer-managed KMS keys and private networking, and keep data and checkpoints in the customer's S3 buckets. AWS says it does not use that data or the resulting weights to train other models.
The launch puts AWS in a niche that smaller players opened first. VentureBeat points to Thinking Machines Lab's Tinker as a closely related service that also separates training logic from GPU management and prices input, sampling and training separately. AWS's pitch is integration with the identity, encryption and storage controls enterprises already run on its cloud.
Several practical questions remain open. AWS's launch post gives no dollar rates or sample training bills and points readers to its pricing page, so it is not yet clear whether per-token billing beats renting dedicated instances for a given workload. And as VentureBeat pointed out, writing the loop also hands customers responsibility for deciding what the model should learn and how to grade it.