Start with the architecture. Then get into the memory math, the software, and the choices that make a homelab useful.
A Windows and Ubuntu setup across NVIDIA and Intel. Separate model servers, a single front door, VRAM estimates, and Tailscale-only remote access.
What the weight file tells you, what the KV cache adds, and why “it loads” is different from “it fits comfortably.”
Connecting the serving stack, choosing an endpoint, and keeping remote access private.
VRAM capacity, backend compatibility, power, and the tradeoffs worth checking before you buy.
New guides, useful experiments, and the occasional hard lesson.