Your own AI assistant — De-clouding

Layers this moves to you: DataModelCompute

What you gain

Everything you type into a hosted assistant — the medical question, the contract you pasted, the message you were nervous about sending — is stored by the provider and, depending on the terms you accepted, may be used to improve their models. A local model runs on your own hardware. The conversation happens in your house and ends there.

You also get an assistant that reads your documents, notes and photos without uploading them anywhere, works without an internet connection, and never changes its behavior because of a pricing decision in someone’s boardroom.

Be honest about the trade

Local models are smaller than the hosted frontier ones. For everyday use — writing, summarizing, answering questions, helping with code, searching your own files — a good 8–14 billion-parameter model is already excellent. For the hardest reasoning tasks, the hosted giants still win. Most people end up with both: local for daily life and anything private, hosted for the occasional heavy lift. That’s a fine place to be.

The tools (all free)

Ollama or LM Studio — one install, pick a model, start chatting. LM Studio has a friendly window; Ollama is a tiny background service other apps talk to.
Open WebUI — a chat website that runs on your box, talks to Ollama, and gives you the familiar chat experience with your own documents attached. Works from your phone over your private mesh.
Home Assistant can use your local model as the brain behind voice commands.

Which model

This changes monthly, so ask the assistant (prompt below) to check current options. The shape of the answer: with 8 GB of memory, a 3–4 billion-parameter model; with 16 GB, 7–8 billion; with 32 GB or a Mac with shared memory, 12–14 billion and up; with a GPU of 16–24 GB, 20–30 billion at speed. Prefer models with permissive, open licenses.

What you still rent

A hosted assistant for the hard problems, if you want one — ideally under a paid plan whose terms say your conversations aren’t used for training. Read that line before you rely on it.

What this looks like at each budget

Tier 0, $0: LM Studio or Ollama on the laptop you own. A 4–8B model answers in a few seconds. Nothing leaves the room.
Tier 1, one small box: Ollama on the small box so it’s always available, slowly, to every device — or the box acts as the front door to your laptop when it’s on.
Tier 2, a real home server: A Mac mini or a 32–64 GB mini PC running a 12–14B model at comfortable speed, Open WebUI for the family, your documents indexed locally.
Tier 3, the hobby got serious: A discrete GPU, 20–30B models at conversational speed, image and voice models alongside. This is where the hobby lives.

Hand it to Claude

Fill in the brackets, paste the whole thing into Claude (or any assistant you trust), and let it walk you through the rest one step at a time.

I want to run an AI assistant on my own hardware instead of a hosted one, mainly for [writing / summarizing / answering questions / helping with code / searching my own documents / smart-home voice]. Hardware I have or plan to have: [computer model, memory, and graphics card if any]. Budget tier: [0 / 1 / 2 / 3 — see de-clouding.com/tiers]. Comfort: [none / some / fluent]. Operating system: [macOS / Windows / Linux].

Check the current landscape (model releases change monthly) and recommend: one easy way to run models on my machine, and two or three specific open-weight models that fit my memory, with their license and what each is good at. Be realistic about speed and about where a hosted model would still do better. Then walk me through installing, downloading a model, having my first conversation, and — if I asked for it — attaching my own documents or connecting to my smart home, one step at a time, waiting for my confirmation each time. Nothing I type should leave my machine; point out any setting that would change that.