Written 9 Sep 2026. This is a how-to, not a product card.

A GPU can wait

A 1B model on the leftover-box CPU is the first try. An RTX card waits.

Local chat is the last weekend-path step: run a local chat on the old PC. CPU is fine. A used GPU, a new N100 with “AI,” or a Mac Studio does not make Pi-hole more true. One leftover box is enough: one leftover box.

What you already have

About 8 GB of RAM and a few free gigabytes on disk. Peek first: how much memory and how much room. Pull llama3.2:1b. Chat over SSH. Don't forward port 11434.

What next

Weekend path: your first weekend. Tick Lab plan after a real reply. A hypervisor still waits: a hypervisor can wait.

Optional: leftover-box CPU vs a later GPU as a sandbox in ByteRush — same Labthrift passkey as Plan. Skip if you just want to keep reading.

Skip

Skip an RTX, a used mining card, and a Mac Studio this weekend.
Skip a new N100 because the leftover box “can't do AI.”
Skip forwarding 11434 so a phone can chat from the café.

Sources

Walk: local chat on the leftover box. Memory peek: how much memory. Lessons: ByteRush.

Also on this topic

Run a local chat on the old PCOllama plus a small model over SSH. CPU is fine. See how much memory the computer already hasBefore you buy a tiny computer, check RAM. One leftover box is enoughPi-hole, the folder, and SSH share one computer.