Local LLM PC UK — Run AI Models On Your Own Machine

Running a large language model on your own machine keeps your data private, costs nothing per question, and works with no internet connection. This page explains what hardware that actually needs, and which of our UK-built PCs fit — including the ones that do not.

In short

• Video memory (VRAM) is the limit that matters. It decides the size of model you can run, far more than the processor does.
• 12GB is our sensible starting point, 16GB is the value sweet spot, and 32GB is the largest we fit.
• NVIDIA is the easier road. Almost all local AI software expects CUDA. AMD works but needs more setup.
• System RAM is your safety net when a model will not fit on the card — slower, but it opens models that would otherwise fail.
• We cannot serve everyone. Full 70B work needs more VRAM than a consumer card has, and we will tell you that rather than sell you the wrong machine.

How much VRAM do you need to run an LLM?

VRAM is the memory built into the graphics card, and it is the one number that decides which models will load. A model has to fit into it whole, or performance falls off a cliff. The figures below are for 4-bit quantised models, which is what most people run at home, and they leave room for a normal context window.

Which graphics card should you choose?

Pick on VRAM first and brand second. Almost every local AI tool is written for NVIDIA CUDA, so an NVIDIA card gets you running with less effort. AMD cards give you more memory for the money and work well once set up, which suits people who do not mind the extra step.

• RTX 3060 12GB — builds from £1,096. The cheapest card we sell that we would recommend for AI. 12GB and CUDA support.
• RTX 5060 Ti 16GB — builds from £1,418. Our value pick. 16GB of VRAM opens up 24B models.
• RX 7900 XT 20GB — builds from £1,634. The most VRAM per pound on the site, if you are happy with AMD ROCm.
• RTX 5070 Ti 16GB — builds from £1,784. Faster than the 5060 Ti with the same 16GB, if speed matters more than capacity.
• RTX 5090 32GB — builds from £4,712. The most VRAM we fit, and the only card here that handles 32B models with a long context.

Does system RAM matter for local AI?

Yes, as a fallback rather than a first choice. When a model is too large for your graphics card, tools such as Ollama and LM Studio can move part of it into system RAM. It runs much more slowly than the card on its own, but it is the difference between a model opening and a model refusing to load.

Most of our machines take up to 192GB of DDR5 — that covers AMD Ryzen 8000 and 9000, Intel Core Ultra and Intel 14th generation builds. Our AMD Ryzen 5000 machines use DDR4 and stop at 128GB, and some builds cap at 128GB DDR5. The configurator shows the real ceiling for whichever machine you pick, so nothing is guesswork.

What we honestly cannot do

We would rather lose a sale than sell you a machine that will not do your job. There are three things worth knowing before you buy, and none of them are hidden further down a specification sheet.

• Our ceiling is 32GB of VRAM. A 70B model at good quality needs around 40GB. It does not fit, and no amount of tuning changes that.
• Every build uses a single graphics card. We do not fit two cards to add their memory together.
• We do not sell professional AI accelerators. No RTX 6000, no A100, no H100. If your work needs one, a consumer PC is the wrong answer.

If any of that rules us out, tell us what you are trying to run and we will say so plainly rather than take the order.

Our recommended local AI builds

Every machine below is built and tested in the UK, comes with a 3-year warranty, and is priced £100 below our own eBay listing. You can change any part in the configurator, so treat these as starting points rather than fixed packages.

• Getting started — from £1,096. RTX 3060 12GB. Comfortable with models up to about 14B.
• Best value — from £1,418. RTX 5060 Ti 16GB. Handles 24B models and most day-to-day local AI work.
• Most memory for the money — from £1,634. RX 7900 XT 20GB. Reaches 32B models if you are comfortable with AMD.
• No compromise — from £4,712. RTX 5090 32GB. The largest models and longest context windows we can support.

Browse every build or configure your own from scratch.

Frequently Asked Questions

How much VRAM do I need to run a local LLM?

VRAM is the single limit that decides which models you can run. As a rough guide at 4-bit quantisation, an 8B model needs about 8GB, a 14B model about 12GB, a 24B model about 16GB, and a 32B model about 20GB. Leave 2–3GB spare for your context window.

Can I run a 70B model on one of your PCs?

Not comfortably. A 70B model at 4-bit needs roughly 40GB of VRAM, and our largest card is the RTX 5090 32GB. You can load a 70B at heavier 3-bit compression, but quality drops and there is little room left for context. For full 70B work you need more VRAM than we fit.

Which graphics card is the best value for local AI?

The RTX 5060 Ti 16GB is the value pick, from £1,418. It gives you 16GB of VRAM and NVIDIA CUDA support, which almost every local AI tool expects. If your budget is tighter, the RTX 3060 12GB from £1,096 still runs 14B models well.

Is AMD or Intel any good for running AI models?

They work, but with more effort. Most local AI software is built for NVIDIA CUDA first, and AMD ROCm support on Windows is still behind. The RX 7900 XT 20GB from £1,634 gives you the most VRAM per pound we sell, so it suits people happy to do some setup work.

Does system RAM matter, or only the graphics card?

The graphics card matters most, but system RAM is your fallback. When a model will not fit in VRAM, your software can move part of it into system RAM. That runs far slower than the card alone, yet it lets you load models that would otherwise not open at all.

How much system RAM can I have?

Most of our builds take up to 192GB of DDR5. That covers AMD Ryzen 8000 and 9000, Intel Core Ultra, and Intel 14th generation machines. Our AMD Ryzen 5000 builds use DDR4 and stop at 128GB, and some other builds cap at 128GB DDR5, so check the machine you are configuring.

Do you sell professional AI cards like the RTX 6000 or A100?

No. We build with consumer graphics cards only, and every build uses a single card rather than two. That keeps our machines quieter, cheaper and simpler to support. If your work genuinely needs 48GB or more of VRAM, a consumer build is the wrong tool and we will say so.

What is the cheapest PC that runs local AI properly?

A build with the RTX 3060 12GB from £1,096. We would not go below 12GB of VRAM for AI work. Cheaper 8GB cards run small 7B and 8B models only, and you will hit the ceiling quickly once you try anything larger or use a long context window.