AWS Opens SageMaker Training Sessions for Custom AI Model Workflows

Amazon Web Services (AWS) made SageMaker AI Training Sessions generally available Thursday, Oct. 22, giving developers a way to write custom model-training routines while SageMaker manages the GPU operations.

The Python interface exposes sample(), forward_backward() and optim_step(). Developers can combine these operations for supervised fine-tuning, reinforcement learning or a training objective of their own. SageMaker assigns GPU capacity, coordinates distributed jobs, recovers interrupted work and saves checkpoints. AWS also released a cookbook with ready-to-run examples for supervised fine-tuning and reinforcement learning.

Four base models are supported at launch: gpt-oss-20b, gemma-4-31B-it, Qwen3.6-27B and Qwen3.8-27B. Training Sessions are available in US East (N. Virginia), US West (Oregon) and Asia Pacific (Mumbai). Model access differs by location; Qwen3.8-27B, for example, is listed in US West (Oregon).

Billing is based on processed tokens during prefill, sampling and training. Reward scoring performed in a customer’s own environment does not incur GPU-time charges.

Karini AI CEO Nitin Wagh said his team fine-tuned gpt-oss-20b for multi-turn tool calling against an Amazon Neptune knowledge graph. “Setup just worked, with no manual tuning,” Wagh said.

The product sits between preset serverless customization and infrastructure-focused options such as SageMaker Training Jobs and HyperPod. Developers can set the post-training procedure while AWS continues to run the GPU layer.

AWS introduced serverless customization in December 2025, added full fine-tuning for more than 25 open-source models in August and announced configurable sequence lengths on Oct. 21. The new APIs add another way to customize models within SageMaker.

The release follows Together AI’s Oct. 19 post-training launch, which offers reserved GPU capacity for model customization.