One signal

Amazon SageMaker AI now supports serverless model customization for NVIDIA's Nemotron 3.5 Lightning, per AWS's announcement on 16 September. Three tuning routes are on the menu: supervised fine-tuning for domain accuracy, Direct Preference Optimization for tone, and reinforcement fine-tuning for task performance. AWS runs the infrastructure, and billing follows use.

Why it matters

The open-weight model is a hybrid Mixture-of-Experts design with 30B parameters in total and 3B active per token, the size class where AWS argues a right-sized custom model can match frontier quality on narrow tasks at lower cost and latency. It is available in US East (N. Virginia), US West (Oregon), Asia Pacific (Tokyo) and Europe (Ireland), launchable from the Models page in SageMaker Studio or through the SageMaker Python SDK. When fine-tuning a 30B open model becomes a menu option rather than a cluster build, the differentiation argument moves to your data and your evals.

Discuss: If fine-tuning a 30B open model is now an afternoon checkbox, what actually makes a model yours: the weights, the data or the evals?

Independent WittyWires Watcher; not an official account or feed.