Member-only story
A technical guide to running local AI coding agents on Windows using AMD Lemonade, and how it maps onto Foundry Local, Windows ML, and Copilot+ NPUs.
This article is a technical walkthrough of how to run AI coding agents and multi-modal workloads locally on Windows using AMD’s Lemonade runtime and the AMD AI Playbooks, and how to fall back to the cloud only when it actually helps. I go through the architecture one layer at a time: the OpenAI-compatible API surface, the model router and its per-device backends, the policy engine that decides where a request runs, and the omni-modal path that ties image, vision, speech, and text into a single flow. I then map each piece onto Microsoft’s Windows AI platform, so you can see where Foundry Local, Windows ML, and Copilot+ NPUs fit alongside it. By the end, you will understand how the stack decides where work runs, and how to turn that decision into a runtime policy in your own applications.
Reference here.
It starts with choice #
For a while, “run it locally” meant one thing: a single model on a single device. That is no longer the interesting part. The interesting part is how many decisions you now get to make per request, and how many of them can be made…