What hardware do you need to run AI locally?
By Savior · 02 Nov 2026 · Updated 06 Nov 2026
The short answer
It comes down to memory. Compressed to the usual size, a small model (7 to 8 billion parameters) needs about 4 to 5 GB, a mid-sized one (around 30 billion) 16 to 24 GB, and a large 70 billion one about 40 to 48 GB. For most small businesses, a Mac Studio from £2,499 or an NVIDIA DGX Spark at about £5,249 is enough. Graphics cards for PCs are selling at several times their list price, so this is a bad time to build one.
Memory is what matters
An AI model has to fit in the machine's memory to run well. Models are described by their size in parameters ("8B" is 8 billion), and they're usually compressed (quantised) to about 4 bits per parameter, which keeps most of the quality at a fraction of the size. Roughly:
- 7 to 8B: about 4 to 5 GB. Fine for drafting, short summaries and simple questions
- 14B: about 8 to 12 GB
- 30 to 32B: about 16 to 24 GB. The sweet spot for most business work
- 70B: about 40 to 48 GB. The largest most small businesses would run
You need some memory on top for the documents you give it, so leave headroom.
The machines
These are UK starting prices we checked in October 2026. They change, so check before you buy.
- Mac Studio, M5 Max: from £2,499 (Apple UK). The starting model has 36 GB of memory, enough for mid-sized models. Quiet, small and uses little power.
- Mac Studio, M5 Ultra with 96 GB: £5,499 (Apple UK). Room for 70B models with space to spare.
- NVIDIA DGX Spark: £5,249 at Scan, £6,390 from NVIDIA's UK store. A small box with 128 GB of memory, built for running AI.
- A PC with an NVIDIA graphics card. The fastest option, but not now: the RTX 5090 lists at £1,799.99, yet in late October 2026 the cheapest at Overclockers UK was £6,989.99, as demand for AI hardware outstrips supply.
Which one for a small business?
For a small team taking turns, a Mac Studio is where we'd look first: it runs mid-sized models comfortably, it's quiet enough for an office and it costs little to run. If you need larger models or more people using it at once, a Mac Studio with more memory or a DGX Spark is the next step. We'd rather you bought the smaller machine and upgraded than paid for power you never use.
The software
The models are free to download. Popular tools for running them include Ollama, LM Studio and llama.cpp, with a chat window such as Open WebUI or AnythingLLM on top so your team can use it in a browser.
Check each model's licence before you rely on it at work. Several popular families, including Qwen, gpt-oss, Mistral and Phi, use permissive licences such as Apache 2.0 or MIT. Meta's Llama 4 has its own licence with extra conditions, so read it first.
Our own test machine
We're building our own machine now, and we'll publish the speeds we actually measure on it, model by model, so you can see what to expect before you buy.
Next steps
Work out whether it pays for itself with the local vs cloud AI cost calculator, read what local AI really costs compared with subscriptions, or see what we set up in Private AI.