AWS makes training sessions generally available for SageMaker AI

AWS announced on October 22 that training sessions are now generally available for Amazon SageMaker AI. The feature lets developers control how an existing model learns while AWS manages GPU compute resources, simplifying fine-tuning for enterprise AI agents.

Training sessions combine granular programmatic control with a serverless approach. Developers direct operations such as response generation, update calculation, and application through API calls, while AWS handles GPU execution. Both Low-Rank Adaptation (LoRA) and full fine-tuning are supported, allowing teams to define custom learning objectives and combine training stages.

AWS notes that the capability extends its serverless customization offering rather than introducing fine-tuning to SageMaker for the first time. Developers can pause to inspect results and adjust procedures, and reinforcement learning experiments can run without a separately deployed inference endpoint.

Billing is based on processed tokens rather than reserved GPU capacity, and the meter pauses while customer scoring code evaluates a batch. AWS confirmed support for gpt-oss-20b, gemma-4-31B-it, Qwen3.6-27B, and Qwen3.8-27B, with a cookbook covering chat, math, programming, and preference learning. It also provides agent skills for coding assistants including Kiro, Claude Code, Codex, and GitHub Copilot.

Early adopters include ImagineSoftware, which used the service for healthcare claim denial triage, and Karini AI, which customized gpt-oss-20b for a knowledge-graph agent. AWS has not disclosed specific pricing or quantified benefits, directing customers to its SageMaker pricing page.