AI PCs and local LLMs

Running a language model on your own machine is a memory problem before it is a speed problem. A model that does not fit in your VRAM either spills into system RAM and crawls, or refuses to load at all — which is why every guide below starts from the same question: what actually fits? Once that is settled, the rest is choosing between three very different kinds of hardware, and knowing what each is genuinely fast at.

Start here: what fits in your VRAM

Before any buying decision, work out which models your card can hold at which quantisation. Our calculator does the arithmetic — parameters, precision, context length and KV cache — and tells you what runs comfortably rather than what technically loads.

→ LLM VRAM calculator — model size vs card, with the quantisation trade-off spelled out.

→ How to run LLMs locally: the 2026 hardware guide — the software stack and what each tier of hardware gets you.

Which machine to buy

The honest answer depends on whether you want the fastest tokens per second, the most memory per dollar, or something quiet that sits on a desk. Those three goals point at three different machines, and the comparison below is the one most buyers actually need before they spend anything.

→ The best AI PC for running LLMs locally in 2026 — picks by budget, with the memory ceiling and bandwidth stated for each.

→ Mac vs PC for running local LLMs — unified memory versus discrete VRAM, and who each one suits.

→ DGX Spark vs Strix Halo vs Mac Studio — three routes to a lot of addressable memory, compared directly.

The parts that matter

Inference is dominated by memory bandwidth and capacity, so the CPU matters far less than most build guides imply — but it is not irrelevant, and it decides how much you can offload when a model will not fit on the card.

→ Best CPU for AI and local LLMs — where the CPU actually becomes the bottleneck.

The measured numbers

Claims about local-LLM speed are easy to make and hard to check. Our benchmark table puts the silicon, memory, memory bandwidth and quantisation next to every tokens-per-second figure, because throughput without the conditions that produced it is not a measurement.

→ AI PC benchmark: local-LLM speed and memory — what each machine does, under conditions you can reproduce.

How we choose

We synthesise independent testing and vendor specifications, state the conditions behind every number we quote, and say plainly when a figure is a manufacturer claim rather than a measurement. Our full method and how our editorial scores work is on the methodology page; funding is covered in our affiliate disclosure.