Mini-PCs in the AI Test: Apple vs. AMD and Intel with Local LLMs

TVS News AI, editorially supervised, presents Mini-PCs in the AI Test: Apple vs. AMD and Intel with Local LLMs.

A desk full of mini-PCs and an automated benchmark for local Large Language Models — at Golem, editor Mike Faust built exactly this setup and let three platforms compete against each other: Apple, AMD, and Intel. If you're wondering whether a compact machine on your desk can really be a serious AI workstation, this test provides a solid foundation.

The clever part: instead of manually feeding individual apps, the comparison runs automatically. That makes the measurements reproducible — and strips the marketing polish off manufacturer claims. You get numbers that show how many tokens per second actually come out of the machines when a language model runs directly on the device.

The topic hits a nerve. Anyone who wants to run an LLM locally quickly ends up at the classic gaming PC with a fat GPU. But not everyone has the space, power budget, or desire for a noisy tower. Mini-PCs are the obvious alternative — and the selection in 2026 is bigger than ever.

What Golem's test does exactly

The test relies on an automated benchmark that puts local LLMs through their paces on the mini-PCs. Golem editor Mike Faust covers PC & Notebooks as well as Mac at Golem — so he brings experience from both camps. That matters because the Apple ecosystem (Unified Memory, Neural Engine) follows entirely different rules than x86 with a dedicated NPU.

Three platforms are in the ring:

Apple relies on Unified Memory: RAM and GPU user memory share one large pool. That's an architectural advantage for AI, because models load directly into the already generous working memory. The Mac mini shows up in multiple current buyer's guides as the reference for compact local AI.

AMD is rolling out its own NPU line with Ryzen AI — the current Strix Halo generation (Ryzen AI Max+ 395) is being marketed in comparison tests especially for maximum AI flexibility. This is the platform if you want to run 70B models or larger locally.

Intel brings NPUs to the mini-PC class as well with Lunar Lake and Core Ultra, but positions itself differently — more as an all-rounder with solid AI acceleration rather than a pure AI heavyweight. Geekom and other manufacturers rely on Intel's platform for their current models.

Why mini-PCs for local AI are interesting right now

If you want to use AI without sending your data to the cloud, there's no way around local hardware. The classic recommendation is: an RTX GPU with as much VRAM as possible. But that's where the problem starts: an RTX 4090 consumes significant power at load, takes up space, gets loud, and costs as much as a decent used car.

Mini-PCs take a different path. They're small, quiet, use less power, and can still run 24/7 — for example as a home server that serves up a language model via Ollama. Models from 7B to 70B cover today's common use cases, and the form factor fits next to the monitor or under the desk.

The price range cited in current buyer's guides runs from around 490 euros for an entry-level mini-PC with an NPU up to workstations that can handle 70B models. That's a different league than a dedicated GPU tower.

What you can take away from the Golem test

Even though the concrete benchmark numbers come from the full Golem article, you can already deduce quite a bit from the test setup itself: if you want a fair comparison, you have to think about platform differences. Apple's Unified Memory allows different model sizes than Intel's or AMD's dedicated NPUs. AMD's Strix Halo goes in a similar direction to Apple with a lot of shared memory, but positions itself for x86 users.

In practical terms, that means: if you're already in the Apple ecosystem and want to experiment with models up to 70B, a Mac mini with enough RAM is worth considering. If you're on x86, want as much AI performance per watt as possible, or prefer Linux/open-source toolchains, AMD Strix Halo is the most exciting platform in 2026. Intel scores if you're looking for an unobtrusive all-rounder that also handles Office and a browser without breaking a sweat.

Bottom line: the test delivers what ad promises can't

The value of Golem's test lies less in the table than in the setup. An automated benchmark means you can trust the result, because nobody helped manually. And because three architectures with completely different strengths go head-to-head, you also learn something about the platforms — not just about a single device.

If you're seriously considering a mini-PC for local AI, the full article at Golem is required reading. But plan time for your own due diligence: Which models do you want to run? What token rate do you need for your use case? What ports does your setup need? No benchmark answers those questions for you.