The number one question we get from businesses considering local AI: "do I need to buy a monster PC?". Short answer: no. The models that handle everyday office work — drafting, summarising, answering questions over your documents — run comfortably on mid-range hardware. Here's what actually matters.
The one number that matters: RAM
A language model must fit in memory to run. That makes RAM (or video memory, if you have a discrete GPU) the deciding factor — far ahead of CPU speed. As a rule of thumb: 8 GB runs only small models and leaves no room for anything else; 16 GB runs capable 7–8B models comfortably; 32 GB runs the 12–14B class that handles nuanced French very well and lets you keep your other apps open.
Three real budgets
- Around €600 — a mini-PC with 32 GB of RAM (Ryzen-class CPU). No discrete GPU: responses stream at reading speed rather than instantly. Perfectly serviceable for one user; this is the sweet spot for most small businesses.
- Around €1,200 — a desktop with 32 GB RAM and a mid-range GPU with 12–16 GB of VRAM. Responses become near-instant, and heavier models open up. The right choice if AI becomes part of everyone's daily workflow.
- Your existing machine — if your work laptop already has 16 GB of RAM, try before you buy. A 7–8B model may be all you need, and it costs nothing to find out.
Mac or PC?
Apple Silicon Macs are excellent local-AI machines: their unified memory serves as both RAM and VRAM, so a MacBook with 24–32 GB runs mid-size models remarkably well. If your team is already on Mac, you likely don't need to buy anything. On PC, Windows and Linux are equally capable — pick what your team already knows.
One machine or one per person?
You don't need one AI-capable machine per employee. A single capable computer on the office network can serve the whole team's AI requests — that's how we deploy Locaible in most small firms: one well-equipped machine, everyone benefits, and no client data ever leaves the office.
Bottom line: 32 GB of RAM is the comfortable target, €600 is a realistic entry point, and the machine pays for itself once — unlike cloud AI subscriptions billed per seat, per month, forever.