SageMaker AI training sessions are generally available: write your own post-training loop and reward function while AWS runs the GPUs, billed by tokens processed
Reinforcement learning on open-weight models gets easier to start. What you write, what AWS runs, and how it bills.
// Key points
- Training sessions extend SageMaker AI serverless model customization: you write the training loop, custom loss and reward signals, and SageMaker AI handles GPU allocation, distributed execution, fault recovery and checkpoint storage.
- SFT, DPO, RLVR, RLAIF and multi-turn RL are supported, with LoRA or full fine-tuning. The reward function runs in your own environment, where it also connects the agent to its tools or simulators.
- You pay for the tokens SageMaker AI processes; the meter stops while your reward function scores a batch, and there's no capacity to reserve.
- A companion cookbook ships model-customization agent skills that work in SageMaker Studio, Kiro, Claude Code, Codex and GitHub Copilot.
Builder's takeNot paying while your reward function scores matters a lot for multi-turn agent training: a slow evaluation environment no longer burns GPU money. I'd take an end-state reward like the ThinkingBox one in the next item, run a few dozen steps on a small open model to see the curve, and only then decide whether a fixed task deserves its own specialized model.
// Background · from #Open models
Full timeline →- Oct 22 NVIDIA opens Kumo Relational, which predicts churn and fraud directly on relational databases without training and beats supervised pipelines on 41 tasks across 15 databases
- Oct 22 IBM releases Granite Speech 5.0 2B under Apache 2.0, with a 5.06% average English WER and keyword biasing for up to 2,000 terms
- Oct 22 Dharma AI rebuilds a fine-tuned Qwen3-4B's vocabulary around token cost, cutting a production essay grader's measured inference cost 77%