We benchmarked 9 local LLMs on a mini PC

Sixty hours of benchmarks on a $600 mini PC answered the question we keep getting: which local model is actually usable without a discrete GPU?

Why it matters

Quantization quality mattered more than parameter count. A well-quantized mid-size model beat a sloppily converted large one on every task we threw at them, from summarization to code review.

Memory bandwidth, not compute, was the bottleneck on every machine we tested — which is exactly why the new unified-memory mini PCs punch so far above their price.

The bottom line

We benchmarked 9 local LLMs on a mini PC is not an isolated story — it is part of a broader shift we have been tracking across the industry all year. The vendors that adapt their roadmaps early will set the defaults everyone else lives with, and the ones that wait will spend the next product cycle catching up. We will keep this piece updated as the picture changes.

Demo content seeded for layout testing.