AWS lets teams write their own model training loop in SageMaker

AWS announced that training sessions in Amazon SageMaker AI are open to all customers. A team writes its own training loop, with its own loss function (a score of model error) or reward signal. AWS does the GPU allocation, fault recovery and storage of checkpoints (saved model data). Four open models work: gpt-oss-20b, gemma-4-31B-it, Qwen3.6-27B and Qwen3.8-27B. AWS gives no test results.

How it works

Your code operates the loop and the reward function in your own environment. The session makes responses with a sampler, and a trainer updates the model weights. You pay for the tokens. The meter stops while your code scores a batch.

The proof

AWS gives no test results or speed numbers. It gives two customer quotes. ImagineSoftware built a model that sorts healthcare claim denials. Karini AI adjusted gpt-oss-20b for tool use and says setup needed no manual adjustment.

What is not known

The source gives no price for each token. It does not show that a custom loop gives better results than the managed alternatives of AWS. AWS says it does not use your data or your adjusted weights to train a different model.

Insight: AWS wants to own the place where agents are trained

The agent skills for training work in Kiro, Claude Code, Codex and GitHub Copilot. In October 2026 AWS also released open source Strands tools to control agents. This reads as AWS wants to own the training and operation layer. The tool that writes the code can be a different one. The customer keeps the weights, and AWS earns on tokens. Leaders: judge lock in at the GPU layer.