April 21, 2026/599 words/3 min read

The Comfort of Local-First Everything

Running things closer to home feels good!

A while ago I wrote about my accidental homelab. At the time, it was a cheap M1 Mac mini and one of the most useful machines I've ever owned.

That setup has now been replaced by another Mac mini, this time with 64GB of RAM. It is overkill, and I do not regret the upgrade at all. The new machine is even smaller and still quiet enough to forget about. The extra memory means it can run most of my personal infrastructure and a few local models without feeling like I am asking too much of it. The old Mac mini is still around as a dedicated Stardew Valley machine for the missus.

The new machine still runs the databases and background jobs behind my tiny apps. It handles search too, along with my music tracking, bookmark stuff, and the weird automation that would be hard to justify as a standalone product but is incredibly useful to me. None of it needs to be in the cloud. I like knowing my data and weird little workflows, half-finished experiments included, are sitting on a machine I understand and can fix when I inevitably break it.

The newer change is running models locally. I am not pretending a Mac mini is a rack of H100s, but the local scene has gotten good enough that useful models are no longer hard to run!

Gemma 4 just landed and feels tailored for someone like me. The models are released under Apache 2.0 and sized for actual hardware. They have long contexts and multimodal support without feeling like cloud models squeezed down into local shape (gpt-oss). The 26B MoE and 31B dense models are the exciting ones, while the smaller E2B and E4B variants are clearly aimed at edge devices.

That range fits how I use local AI. For classification or cleanup, a tiny, fast model is enough. A document or my weird handwriting sometimes needs more. Most of the time I want to ask something without sending the whole context to a vendor.

Qwen has become hard to ignore, especially Qwen3, because its small models hit above their weight and its larger MoE ones are serious. Phi interests me too because it is tiny enough to make me rethink how much model a task needs.

My default path now starts at home. If a task needs a frontier model, fine, I will use one. I pay for enough of them already! A log does not need frontier intelligence, and neither does a note that only needs metadata. Those jobs can stay home.

Tailscale is the reason this setup feels sane. I know people say "it just works" too often, but Tailscale really does! The Mac mini stays home while my laptop and phone go wherever I do, often through some cursed hotel Wi-Fi. Everything still feels like it is on the same network.

Having that small computer at home changes how I build things. I can make small software around my own data without wondering which tool has some weekly usage left or having to design a privacy policy for myself. I can also run experiments that would never justify a hosted product because they are too specific to me or simply too stupid.

I am not moving everything into a bunker. The cloud is still useful, and I like the balance better now. I still reach for Codex or Claude a lot because there is a great deal of work a local model cannot (yet!) do. When the task is small, it is usually easier to let the Mac mini handle it.