Agentic Local Models

A lot of developers are talking about running AI privately on their own machine, but what does that actually look like when the work you care about lives inside FileMaker? If you handle sensitive client data, or you would simply rather your solution's structure never leave your disk, the promise of a local model is hard to ignore. The question is whether the hardware sitting in front of you can really deliver, and whether the results hold up once you move past the hype and start doing real development work.

In this tutorial you will watch a single FileMaker prompt run two ways, once against a fast cloud model and once entirely on a local machine with nothing touching the internet. You will see the live memory and GPU behavior as the local model loads, the tokens per second it actually produces, and the honest gap in quality between a small local model and a frontier one, right down to a hallucinated function that never should have existed. Along the way you will understand why parameter counts, quantization, dense versus mixture of experts, and prompt processing all shape what you can realistically expect from your own setup.

The video goes on to build a practical picture of an agentic workflow around your FileMaker solution, from feeding whole-solution context to an agent, to choosing hardware that reaches genuinely usable speeds, to running multiple agents at once and managing the context window that makes or breaks the result. You will also see where the platform itself is heading, what it means to move from writing code by hand to directing an agent that writes it for you, and how to keep every part of that pipeline local when privacy is not negotiable. It is the practical grounding that turns a curiosity about local AI into a real decision about your own environment.

Click the title or link to this article to view the video.

View the full article