Running a local LLM is about one thing: enough fast memory to hold the model. On a mini-PC there are now two ways to get it. The old way is a fitting low-profile GPU, ranked by its VRAM - but most mini-PCs have no PCIe slot, so the field is small. The new way is a big-APU box (Ryzen AI Max+ 395, or a Ryzen AI 9 HX 370) that allocates most of its RAM as graphics memory - a 128GB 395 can hand roughly 96GB to the iGPU, enough to fit a 70B locally with no add-in card at all.

So this list ranks both paths by the memory each brings: unified-memory boxes by what they can allocate, GPU boxes by their card VRAM. A discrete GPU is faster per gigabyte, but a unified-memory box is how you fit the biggest models cheaply and quietly. Buy for the model size you actually want to run.

What to look for

Memory is the spec
~8GB runs a 7B, ~12GB a 13B, ~24GB a 33B, ~48GB+ a 70B (all Q4). Everything else is secondary.
Unified-memory boxes
A Ryzen AI Max+ 395 / AI 9 HX 370 allocates much of its RAM as VRAM - the cheapest way to fit a 70B, at lower throughput than a discrete card.
Or a discrete GPU
A fitting low-profile card (the slot supplies ~75W, no external connector) is faster per GB - the A2000-class cards are the natural mini-PC fit.
Big card = DIY
Want a 24GB+ full-size GPU for maximum speed? That is a proper build with a case and PSU, not a mini-PC.

The picks right now

# Pick Fit Price
1 Minisforum AI X1 Top pick
Minisforum AI X1 runs models on 96GB unified memory - 70B comfortably, up to ~120B (Q4).
96 GB unified
~$749 Check price
2 Beelink GTR9
Beelink GTR9 runs models on 96GB unified memory - 70B comfortably, up to ~120B (Q4).
96 GB unified
$3,639
3 Framework Desktop
Framework Desktop runs models on 96GB unified memory - 70B comfortably, up to ~120B (Q4).
96 GB unified
~$1,999 Check price

Ranked from curated specs, same logic as the interactive recommender. Approximate - verify against your exact setup. Prices move; check the retailer.

Other guides