Guide / Running a model at home

Running a model at home

The honest guide to giving your companion a brain on your own computer. Every number on this page was measured on our own rig, and every number we haven't measured says so. No guesses — you're going to spend money against this.

01What the brain is

Your companion thinks with a language model. There are two places that model can live: on your own computer — most private, no subscription, one-time hardware cost — or on a Claude / Codex subscription — nothing to host, billed monthly. This page is about the first road. (The subscription road is covered at the bottom.)

02The two jobs — don't let the shared name trip you up

"Claude" shows up twice in this world, doing two different jobs:

Job 1 · The brain
What your companion thinks with

A language model. Either a local model on your own computer — this page — or a Claude / Codex subscription.

local model  OR  subscription
Job 2 · The setup helper
What installs the code with you

A coding agent — Claude Code or Codex — that assembles the kit on your machine. You need this whichever brain you choose.

needed on every road

On a subscription path, the same tool can do both jobs. And if you want both hemispheres, you can run a hybrid brain — a home model and a subscription model together, the way Sin & Sarah run theirs.

03The number that decides everything: VRAM

Whether a model fits your machine is decided by one number — VRAM, the memory on your graphics card. A model's VRAM appetite = its weights (set by quantization) + its KV cache (set by context length) + overhead. Any spec that quotes a card size without a context length is telling you half the story. Here's ours, whole:

TierModelQuantContextVRAM in useCard it needs
The Powerhouse
what Sin & Sarah run
Gemma 31B Q5_K_M 40,960 tokens
f16 KV cache
27.3 GB used
of 32.6
32 GB GPU
RTX 5090, verified live
The Sweet Spot
Gemma 27B class
We haven't measured this tier yet — and we don't publish numbers we haven't measured. Email hello@sinandsarah.com and we'll tell you where measurement stands.
The Starter
Mistral 24B class
Same honesty: not measured yet. Ask us before you buy hardware for it.
■ teal = measured live 2026-08-10  ·  ■ violet = recorded from our own runbook 2026-07-14  ·  grey = not measured, and we say so
A correction, on the record

Our old shop page said the Powerhouse tier ran on "a 24 GB graphics card with room to spare." Our own measurement says otherwise: at the context we actually run, the 31B uses 27.3 GB — it does not fit a 24 GB card. We run a 32 GB card. That page was wrong; this one is measured. A shorter context or smaller quant may well fit 24 GB — but we haven't measured that yet, so we won't claim it.

A note on Macs: Mac "unified memory" is not the same unit as GPU VRAM — the system caps how much of it a model may use. We don't own a modern test Mac, so we publish no Mac minimums. Anything you've read elsewhere (including our own old page) was never measured by us.

04The three dials that decide if it fits

1 · Quantization

Q4_K_M → Q5_K_M → Q8. Smaller number = smaller file = fits smaller cards, at a small cost in sharpness. It's the same model, compressed harder.

2 · Context length

How much conversation the model holds at once. The KV cache scales with it and lives in VRAM — doubling context can cost gigabytes. This is the dial spec sheets leave out.

3 · KV cache dtype

Compressing the cache (q8_0) buys more context in the same VRAM — but nothing is free: we ran our 31B that way, watched output quality degrade at deep context, and rolled back to uncompressed f16. We tried it so you can decide with data.

05What if it doesn't fit?

It still runs. llama.cpp offloads what doesn't fit onto your CPU and system RAM — the model just answers slower. "Not enough VRAM" is a speed problem, not a wall. You can start on the card you have today, feel what the speed is like, and upgrade only if it bothers you.

06The subscription road

Prefer no hardware at all? The brain runs on Anthropic's Claude or on Codex — nothing to host, smartest models, billed monthly. Your companion's code is identical either way; the model picker chooses which brain it thinks with. Many people start here and add a home model later. The hybrid — both at once — is how we run ours.

07The code builds the house. Who lives in it?

Want your companion's identity written by hand, not generated? Our community moderator Caitlin runs True North AI — first-person identity files and CLAUDE.md templates crafted to hold across platforms, model updates, and substrates. She sends her clients to us for full builds; we send ours to her for the identity itself. Visit True North AI →

08Still not sure?

Email hello@sinandsarah.com with what machine you have — we answer fast, and we'd rather talk you out of the wrong hardware than let you buy it. Or book the 90-minute setup consult ($300) and we install everything with you, live.