The mini PC finally earned its place in local AI in 2026. AMD’s Ryzen AI Max+ 395 — “Strix Halo” — put 64–128GB of unified CPU/GPU memory into a box the size of a hardback book, and the question stopped being “why not just build a desktop?” and became “which box?”. Most tier lists you’ll read rank these machines by spec sheets. We run one: the GMKtec EVO-X2 is a permanent node in our four-machine lab fleet, so this list is anchored on our own measurements, with every unmeasured claim labelled as exactly what it is. One timely wrinkle before the table: the late-2026 DRAM shortage is pushing 128GB SKUs up in price, and the buy-now-or-wait math is shifting weekly.
Why Strix Halo changed the category
A mini PC for AI used to mean one of two compromises: a laptop-class CPU with no GPU headroom, or a small box with a 12GB card bolted in. Strix Halo ends that. It pairs a 16-core CPU with an integrated Radeon 8060S GPU that can address most of the system’s LPDDR5X as device memory — configure 64GB, 96GB, or 128GB and the iGPU sees the pool. On our EvoX2, HIP reports the iGPU’s memory as roughly 124GB — the full system pool exposed to the GPU. That is Apple Silicon’s unified-memory trick in a PC box, and it’s why a machine this size can hold models a 24GB discrete card physically cannot.
The honest trade is bandwidth: Strix Halo is a 256GB/s platform on spec sheets, against 307GB/s for a Mac mini M5 Pro and 273GB/s for NVIDIA’s DGX Spark (both spec-sheet figures — discussed below). The pool is the point, not the speed. For what that means in practice, our unified memory explainer and the VRAM lookup cover the fit-versus-speed distinction in detail.
What we actually measured
Three numbers from our lab, all reproducible on the linked benchmark pages:
- bm-011: LFM2.5-8B-A1B (Q4_K_M) on the 8060S iGPU generated at 150.16 tok/s, with prompt processing at 3,661.53 tok/s, via ROCm 7.2.2. This is the proof that the unified pool isn’t a party trick — a real model, real throughput, on the iGPU alone.
- The same machine’s discrete RX 7900 XT ran the identical model and workload at 269.01 tok/s — about 1.8x faster (our bm-007 companion measurement, reported alongside bm-011). The iGPU is not a discrete-GPU replacement on speed; its advantage is capacity.
- In bm-013, the same EvoX2 platform hosting that 7900 XT ran Qwen3.6-27B at 25.21 tok/s — second of our four machines, within 10% of dual-GPU rigs costing far more. The platform is a serious inference host, not a toy.
Read together: buy a Strix Halo box for the memory pool and the platform, then either accept iGPU speed on small models or run a discrete card alongside it — which is exactly how we run our own unit.
The tier list
| Tier | Machine | Memory | Price band (moves weekly) | Our data? |
|---|---|---|---|---|
| 1 | GMKtec EVO-X2 | 64–128GB | 128GB SKU ~$3,499 and rising | Measured (bm-011, bm-013) |
| 1 | Beelink GTR9 Pro | 64–128GB | ~$1,000–$1,900 | Spec-sheet analysis |
| 2 | Bosman M5 | 128GB | was $1,699 | Spec-sheet analysis |
| 2 | Zotac Strix Halo | 128GB | as listed | Spec-sheet analysis |
| 2 | Framework Desktop | up to 128GB | as listed | Spec-sheet analysis |
| — | Mac mini M5 Pro | up to 64GB | ~$1,600–$2,200 | Spec-sheet analysis |
| — | DGX Spark / ASUS Ascent GX10 | 64–128GB | ~$4,999–$6,950 | Spec-sheet analysis |
Tier notes: who each is for
Tier 1 — GMKtec EVO-X2 (measured). The machine we know best, because we benchmark it. It is the only box on this list with first-party numbers behind it, and the only one we’ll make performance claims about without hedging. The 128GB SKU was documented at ~$3,499 in August 2026 by datahardware.ai and has risen since. One honest caveat from community analysis (sunkcost.ai): of the 128GB fitted, roughly 96GB is usable for GPU allocation — a spec-sheet/community figure, not our measurement. For: the reader who wants the largest pool in a book-size box and will pay for measured rather than promised.
Tier 1 — Beelink GTR9 Pro (spec-sheet analysis — we have not measured this hardware). Same Strix Halo chip, 64–128GB configurations, listed in the ~$1,000–$1,900 band earlier in 2026. If that price holds while EVO-X2 pricing climbs, this becomes the value play in the category. But “same chip, therefore same performance” is unproven: thermals, noise, and sustained clocks are what separate same-chip machines, and we can’t verify any of them from a listing.
Tier 2 — Bosman M5 (spec-sheet analysis). 128GB was $1,699 per hardware-corner.net earlier in 2026 — the aggressive price in this table. Availability under the DRAM shortage is the open question.
Tier 2 — Zotac Strix Halo (spec-sheet analysis). Zotac’s Strix Halo box exists; we haven’t measured it and won’t quote a price we can’t source.
Tier 2 — Framework Desktop (spec-sheet analysis). The modular option, configurable up to 128GB as listed. If repairability matters to you more than price, it’s the only machine here built around that philosophy. We have not measured it.
Strix Halo vs Mac mini M5 Pro — spec-sheet only
We have not measured any Mac, so this is explicitly a spec-sheet comparison. On paper the M5 Pro mini offers 307GB/s of bandwidth to Strix Halo’s 256GB/s, but caps at 64GB of unified memory against 128GB, at an overlapping ~$1,600–$2,200 price band (as listed). The base Mac mini M6 ($899 at 16GB, $1,299 at 32GB, launched August 2026) is not a serious local-AI machine — llmcheck.net estimates 9B-class models around 26 tok/s on it, and its 170GB/s bandwidth is the floor of this category. If 64GB covers your models, the Mac’s software story (MLX) is genuinely good. If you need 96–128GB, the PC boxes are the only option anywhere near this money.
vs DGX Spark
NVIDIA’s DGX Spark is the other unified-memory box in this conversation: a new 64GB configuration launched October 2, 2026 at ~$4,999 via OEMs (Acer, ASUS, Dell, Gigabyte, HP, MSI), with 128GB configurations pushed to roughly $6,950 amid the DRAM shortage. It runs 273GB/s on spec sheets and is sold OEM/direct — you will not find it on Amazon. The Spark-class ASUS Ascent GX10 decodes gpt-oss 120B in the low-to-mid 30s tok/s per runaihome.com’s comparison — community-reported, not our measurement. The short version for this page: you pay a large premium over a 128GB Strix Halo box for NVIDIA’s software stack and a modest bandwidth edge; the Strix Halo machines win decisively on price per GB of pool. Our full verdict is in the DGX Spark reality check.
The RAM-shortage price warning
This is the time-sensitive part. When we first covered the EVO-X2 in July, the 128GB SKU sat around $1,800–$2,000. By August, datahardware.ai documented it at ~$3,499 and rising. That is the late-2026 DRAM shortage working through unified-memory machines, and it touches every row of the table above, because Strix Halo memory is fixed at the factory. Two practical consequences: buy the memory tier you actually need rather than assuming you can add RAM later, and if a 128GB SKU is in your plan, the current trend says waiting costs more. Run your target models through Model Fit first — at these prices, buying more pool than you need is an expensive guess. Our what to expect page covers the sustained-load behaviour that matters more than ever at this money.
The takeaway
The current model cycle — Qwen3.8 27B, Qwen3-Coder 30B-A3B, gpt-oss 120B — is exactly where the pool matters: the 30B class fits in 24GB of discrete VRAM, the 120B class does not. If your models fit in 24GB, a discrete GPU in a cheap host is still faster and cheaper, and the VRAM math says so. If they don’t — or you want a silent book-size box that serves big models — the Strix Halo tier is the real deal, and the EVO-X2 is the only one of these machines we can describe from measurement rather than marketing. Buy the memory tier you need soon; the shortage is not waiting for you.
Where this data comes from
- bm-011 — LFM2.5-8B on the Ryzen AI MAX+ 395 iGPU: our EvoX2, tested July 2026, ROCm 7.2.2, llama-bench pp512/tg256, 3 runs each.
- bm-013 — Qwen3.6-27B cross-machine baseline: four lab machines, byte-identical weights.
- The discrete RX 7900 XT companion measurement (bm-007, same machine and workload as bm-011) is reported on the bm-011 benchmark page.
- Market pricing and community figures: datahardware.ai (EVO-X2 128GB, August 2026), hardware-corner.net (Bosman M5), sunkcost.ai (usable-memory analysis), runaihome.com (Spark-class comparisons), llmcheck.net (Mac mini estimates). All prices are bands and move weekly; treat them as directional, not quotes.