The mini-PC I returned: choosing hardware for a local AI appliance

If you're building an appliance around a local language model, the hardware decision is the product decision. Here's the path PHNTM One took — including the box that went back to the store — and what each step taught us.

The prototype: a cheap x86 mini-PC

The first PHNTM One lives on an inexpensive Intel N95 mini-PC driving a 10.1″ touchscreen. As a proving ground it's been excellent: enough CPU to run a small quantized model, a real Debian system, and cheap enough that mistakes weren't expensive. It's still on my desk, still running, still the reference build.

The box that went back

Wanting a second unit, I bought a similar-looking mini-PC built on the older N5095. It looked equivalent on the spec sheet. It was not.

• No AVX2. Modern CPU inference leans hard on vector instructions; the mainstream llama.cpp-family builds and most performance work assume AVX2. The N5095 doesn't have it. On paper, a minor spec-sheet footnote. In practice, the single most important line on the page for this workload.
• An ACPI interrupt storm. The firmware fired a general-purpose event (GPE6F) continuously, and the kernel obligingly serviced it — burning a core at idle, forever. You can mask a GPE and fight your own firmware, but that's not a foundation for an appliance someone else owns.

It went back. Total cost: a restocking window and a lesson.

Why production is Raspberry Pi 5

For the units we'll actually ship, we settled on the Raspberry Pi 5 (8 GB) with a 10.1″ 1920×1200 touchscreen, speaker, and active cooling. Not because it wins raw benchmarks against every x86 mini-PC — it doesn't — but because it wins as an appliance platform:

• One known board. Every unit is the same hardware with the same firmware — no per-vendor BIOS roulette, no mystery ACPI tables. What we burn in is what you get.
• A massive ecosystem. The Pi has more eyes on its kernel, thermals, and quirks than any small vendor's mini-PC will ever get. When something's odd, someone has seen it.
• Honest thermals in a small enclosure. Predictable, quiet, and well understood with active cooling.
• Reproducibility. Production units flash from a golden-master image, boot, and self-test — the path from "parts on a bench" to "working device" is a pipeline, not an art project. Every build then runs a 72-hour burn-in before we trust it.

Would a bigger box ever make sense?

Maybe — a future, beefier SKU with more RAM could run larger models and is genuinely tempting (x86 with AVX2 this time, we've learned). But the sweet spot for a first product is the machine that's cheap enough to own without ceremony, boring enough to trust, and open enough that you can SSH in and see for yourself. That's the Pi.

PHNTM One is a private AI appliance being built by hand — see the specs or join the waitlist. Questions? Email me — I answer everything myself.