Six Windows mini PCs with a real 40+ TOPS NPU, plus the Mac mini, are worth buying in the US right now. Plenty of others carry an “AI” badge and have a tiny NPU, or none at all. The best mini PC with NPU hardware you will actually use comes down to three things: the NPU’s own TOPS figure (not the platform total on the box), how the RAM is fitted, and whether the software you care about can use the NPU yet.
The other thing that changed this year is memory. A 32GB DDR5 SO-DIMM kit now costs about as much as a barebone mini PC, so a cheap barebone can end up costing more than a fully built Copilot+ box with soldered RAM. We price that in below, and we cover what the NPU actually runs on Windows and Linux, which most AI mini PC roundups skip.
Quick picks
- Best overall: Minisforum AI X1 Pro-470. 55 TOPS NPU, upgradeable RAM, OCuLink, and an XDNA2 NPU that Linux LLM tools support.
- Best value Copilot+ box: GMKtec K13. 47 TOPS, 16GB and an SSD included for about the price of a bare chassis.
- Best for business fleets: ASUS ExpertCenter PN54. Ryzen AI 7 350, 50 TOPS, three-year warranty.
- Smallest Copilot+ PC: MSI Cubi NUC AI+ 3MG. Panther Lake in a 0.51 liter case.
- Best Intel performance: Minisforum M2 Pro. Arc B390 graphics and 10GbE next to a 50 TOPS NPU.
- Biggest NPU: ASUS Ascent QN10. Snapdragon X2 Elite with an 80 TOPS NPU, Windows on Arm only.
- Not on Windows: Mac mini (M6). The Neural Engine route, for people who live in macOS.
How we picked
We have not run these seven boxes on our bench, so nothing here is a lab result of ours. Every spec comes from the maker’s spec page, cross-checked against an independent review or owner reports where they exist. NPU TOPS figures come from AMD, Intel and Qualcomm product pages, not from Amazon titles. Prices and stock were read from Amazon’s API on the day this was written. The Linux package status was checked directly against the Arch, Fedora and Ubuntu repositories.
We left out Strix Halo boxes on purpose. They have a 50 to 55 TOPS NPU too, but people buy them for 128GB of unified memory and the GPU, which our Ryzen AI Max+ 395 mini PC comparison covers. Anything below the 40 TOPS Copilot+ bar is out, including Arrow Lake H and Meteor Lake machines sold as AI PCs.
Best mini PC with NPU: specs compared
| Mini PC | Chip | NPU TOPS | RAM | Notable I/O | Price band |
|---|---|---|---|---|---|
| Minisforum AI X1 Pro-470 | Ryzen AI 9 HX 470 | 55 | DDR5 SO-DIMM, up to 128GB | OCuLink, 2x USB4, 2x 2.5GbE | About $730 barebone |
| GMKtec K13 | Core Ultra 7 256V | 47 | 16GB LPDDR5X, soldered | 2x USB4, 5GbE | About $590 with 16GB/512GB |
| ASUS ExpertCenter PN54 | Ryzen AI 7 350 | 50 | DDR5 SO-DIMM, up to 64GB | USB4, 2x 2.5GbE, 2x DP | About $575 barebone |
| MSI Cubi NUC AI+ 3MG | Core Ultra 5 322 | 46 | DDR5 SO-DIMM, 2 slots | 2x Thunderbolt 4, 2x 2.5GbE | About $470 barebone |
| Minisforum M2 Pro | Core Ultra X7 358H | 50 | 32GB LPDDR5X, soldered | 3x USB4, 10GbE + 2.5GbE | About $1,280, no SSD |
| ASUS Ascent QN10 | Snapdragon X2 Elite | 80 | 32GB LPDDR5X, soldered | 3x USB4, 2.5GbE | About $1,700 |
| Mac mini (M6) | Apple M6 | Not published | 16GB unified, soldered | 3x Thunderbolt, 2.5GbE | About $780 |
1. Minisforum AI X1 Pro-470
The AI X1 Pro-470 is the one we’d buy if the NPU is part of the reason you are shopping. Its Ryzen AI 9 HX 470 has a 55 TOPS NPU, the biggest in any x86 mini PC we found on sale in the US, with Radeon 890M graphics beside it. AMD rates the HX 475 at 60 TOPS, but we found no mini PC shipping it yet.
The chassis does the rest. You get OCuLink for an external GPU, two USB4 ports, dual 2.5GbE, three M.2 slots and SO-DIMM slots that take up to 128GB of DDR5. The NPU is XDNA2, so FastFlowLM and Lemonade can run LLMs on it under Linux, which is not true of older Ryzen chips.
Buy it for: anyone who wants a Copilot+ box today and an upgrade path later. The catch: it is not cheap. The barebone is about $730, and two 16GB SO-DIMMs add another $440 to $590 at current prices. The built 32GB/1TB version is about $1,300, but Tom’s Hardware got a review unit with a single SO-DIMM that ran poorly until it was moved to dual channel. Check the stick count before you trust it.
2. GMKtec K13
The K13 is the value pick because it ships complete. For about $590 you get a Core Ultra 7 256V, 16GB of LPDDR5X and a 512GB SSD, which is roughly what a DDR5 kit alone costs right now.
The Lunar Lake chip has a 47 TOPS NPU and Arc 140V graphics, and it clears every Copilot+ hardware line: 40+ TOPS, 16GB and 256GB of storage. It also has eight logical processors, the minimum Recall asks for. The I/O is better than the price suggests, with two full-function USB4 ports, HDMI 2.1 and a 5GbE port.
Best for: Windows users who want the Copilot+ features at the lowest real cost. Where it falls short: the 16GB is on the chip package and can’t be upgraded, and that is the floor for Copilot+ features. Big-outlet reviews are thin, so we are leaning on owner reports and the spec sheet. Stock was low when we checked.
3. ASUS ExpertCenter PN54
The PN54 is the box an IT team can standardize on. ASUS sells it as a Copilot+ PC in its configured versions, backs it with a three-year warranty, and the barebone with the Ryzen AI 7 350 is about $575.
The Krackan Point chip has the same 50 TOPS XDNA2 NPU as AMD’s flagship 300 series, paired with smaller Radeon 860M graphics. ASUS lists two DDR5-5600 SO-DIMM slots up to 64GB, one USB4 port, two DisplayPort outputs, HDMI 2.1, dual 2.5GbE and Wi-Fi 7. Notebookcheck found it strong in everyday office work.
Good fit: offices that want four-display support and a known support channel. Look elsewhere if you game or need an eGPU: the 860M is modest and there is no OCuLink. Owners report that 24Gbit (12GB and 24GB) SO-DIMMs aren’t supported, so stick to 16GB or 32GB modules. It has only a handful of Amazon ratings so far. On the listing, confirm the style reads Ryzen AI 7, since the bullets mix in AI 5 copy.
4. MSI Cubi NUC AI+ 3MG
MSI’s 3MG fits a Panther Lake Core Ultra 5 322 and a 46 TOPS NPU into a 0.51 liter case. MSI is one of the few mini PC brands that calls its box a Copilot+ PC on the listing itself.
It has two Thunderbolt 4 ports (one takes power in), two 2.5GbE ports and two HDMI outputs, plus Wi-Fi 7 and a fingerprint reader in the power button. Tom’s Hardware tested the Core Ultra 9 386H version and found it quiet and easy to open for RAM and SSD swaps. It also called the ports slightly stingy (no rear USB-A, no audio jack) and the graphics weak, and the 322 has an even smaller GPU.
Made for: a desk or VESA mount where size matters more than graphics. The trap: the sticker price is not the real price. The $470 barebone ships with no RAM, no SSD and no Windows license, so it lands well over $1,000 once it is usable.
5. Minisforum M2 Pro
The M2 Pro is the Intel box to get if you want the graphics too. Its Core Ultra X7 358H has a 50 TOPS NPU and the 12 Xe-core Arc B390, the fastest integrated GPU Intel ships right now.
It is also the only pick here with 10GbE built in, next to a second 2.5GbE port, three USB4 ports and three M.2 slots. Memory is 32GB of soldered LPDDR5X, because Panther Lake chips with the big Arc GPU don’t support SO-DIMMs. Intel’s Linux NPU driver lists Panther Lake as supported.
Pick it for: a homelab node or creator box that needs fast networking and a strong iGPU. Pass on it if you need more than 32GB, or you want the NPU for Frigate today. The stock Frigate image bundles an Intel NPU driver older than the release Panther Lake needs. The roughly $1,280 listing ships without an SSD, and Amazon showed a later delivery date rather than stock on hand, so check the estimate.
6. ASUS Ascent QN10
The QN10 has the biggest NPU in any mini PC we found: 80 TOPS on Qualcomm’s Snapdragon X2 Elite, with 18 CPU cores and Adreno X2-90 graphics in a 0.7 liter case.
Windows on Arm gets the longest Copilot+ feature list, because a few Photos and Paint features are Snapdragon only. The Amazon listing is the 32GB LPDDR5X model with three USB4 ports, 2.5GbE and Wi-Fi 7. StorageReview found the fan stops completely at idle.
Ideal buyer: a Windows user who wants every Copilot+ feature and the fastest NPU. Not for you if you run Linux, need x86 drivers, or play games with anti-cheat. Most x86 apps run through Windows on Arm emulation, but not all. Qualcomm’s X2 Linux support is a developer preview, and NPU support on Linux is still being built. At about $1,700 it is also the priciest pick, and 32GB is the ceiling.
7. Mac mini (M6)
The Mac mini belongs on the list because its Neural Engine is an NPU in all but name. It is the easy answer if you do not need Windows or Linux.
Apple doesn’t publish a TOPS figure for the M6, so you cannot compare it with the others on that number. The base model has 16GB of unified memory, three Thunderbolt ports, HDMI and 2.5GbE, and Amazon had it at about $780 against Apple’s $899 list price.
Right for: macOS users who want Apple Intelligence on a desk. Wrong box if you want Copilot+ features or Linux, which this box does not do. The base price is up $300 on the M4’s $599 launch price, and the older model is only sold used or renewed on Amazon now.
AI mini PCs that have no real NPU
This is where most buyers get caught. Several popular “AI” mini PCs use AMD’s Ryzen 7 H 255, which has no NPU at all. Intel’s Arrow Lake H chips have one, but at 13 TOPS it is far below the Copilot+ bar. Big numbers like 86, 97 or 99 TOPS in a listing title are platform TOPS, which add the CPU, GPU and NPU together.
| Model sold as AI | Chip | NPU TOPS | Copilot+? |
|---|---|---|---|
| Beelink SER9 Pro AI | Ryzen 7 H 255 | None | No |
| GMKtec K12 | Ryzen 7 H 255 | None | No |
| Minisforum AI X1-255 | Ryzen 7 255 | None (disabled) | No |
| ASUS NUC 15 Pro | Core Ultra 7 255H | 13 | No |
| ASUS NUC 14 Pro | Core Ultra 7 155H | 11 | No |
The naming trap goes further. Notebookcheck found the Beelink SER9 Pro is slower than the plain SER9, which uses a Ryzen AI chip with a 50 TOPS NPU. Read the processor model, then check its NPU figure on the AMD or Intel spec page.
What the NPU actually runs on a mini PC
An NPU is a low-power engine for small, always-on AI jobs. It isn’t a faster GPU, and for LLMs the integrated graphics match or beat it on speed.
On Windows
Microsoft’s Copilot+ PC requirements are a 40+ TOPS NPU, 16GB of RAM and 256GB of storage. A box with the right chip but 8GB of RAM does not qualify. On a qualifying machine the NPU handles Click to Do, the improved Windows search, Live Captions translation and the local image features in Paint and Photos. Windows Studio Effects (background blur, eye contact, auto framing) also runs on it, and Copilot+ PCs can now apply those effects to a USB webcam.
Recall is the awkward one on a desktop. It needs Windows Hello Enhanced Sign-in Security with a biometric, and external fingerprint readers only became eligible with a July 2026 update. It also needs a Secured-core PC, which comes down to each maker’s firmware. Of our picks, the AI X1 Pro-470, the MSI 3MG and the M2 Pro have a fingerprint reader built in; the K13, the PN54 barebone and the QN10 do not. Microsoft’s own Recall support page now says “Windows 11 AI+ PC” where it used to say Copilot+ PC, with the same hardware bar.
On Linux
Linux support splits by vendor. AMD’s NPU driver has been in the kernel since 6.14, and the useful LLM tooling (FastFlowLM and Lemonade) needs kernel 7.0 (or the amdxdna-dkms backport) plus NPU firmware 1.1 or later. It works on XDNA2 chips only (Ryzen AI 300, 400 and Max), so Ryzen 7040, 8040 and 200 series NPUs are left out. Intel’s NPU driver has been mainline since 6.3 and works through OpenVINO. Qualcomm has no usable NPU path on Linux yet.
What the main distros ship, checked against their repositories on October 4, 2026:
| Distro | AMD XDNA2 stack | Intel NPU stack |
|---|---|---|
| Arch Linux (extra) | xrt 2.21.75, xrt-plugin-amdxdna, fastflowlm 1.0.7, lemonade-server | intel-npu-driver 1.38.0, openvino-intel-npu-plugin 2026.4.1 |
| Fedora 44 | No XRT in the official repos (COPR only) | intel-npu-driver 1.32.0 |
| Ubuntu 26.04 | libxrt-npu2 in universe; in-kernel amdxdna; FastFlowLM from GitHub or installed by Lemonade | No intel-level-zero-npu in the archive; debs from Intel’s GitHub |
Arch is the easiest distro for NPU work right now. On any distro, these commands show whether the kernel found the NPU:
lspci -nn | grep -iE '1022:17f0|8086:(643e|b03e)'
ls -l /dev/accel/
sudo dmesg | grep -iE 'amdxdna|intel_vpu'A working NPU shows up as /dev/accel/accel0, and dmesg shows the amdxdna or intel_vpu driver loading its firmware. PCI ID 1022:17f0 is an AMD XDNA2 NPU; 643e is Lunar Lake and b03e is Panther Lake. AMD’s amdxdna kernel documentation covers the driver side, and on Intel the OpenVINO device list should include NPU.
For local LLMs
Ollama has no NPU backend, so a model pulled with the standard Ollama setup runs on the CPU or iGPU whatever NPU you buy. llama.cpp now has a preview OpenVINO backend that can target the Intel NPU. On AMD, NPU inference goes through FastFlowLM or Ryzen AI Software instead.
Published numbers show the NPU is not a speed upgrade for LLMs. FastFlowLM’s own benchmark has Llama 3.1 8B decoding at about 12.8 tokens per second on the Ryzen AI 7 350’s NPU. A separate test of the bigger Radeon 890M iGPU (in a Ryzen AI 9 HX 370) with Vulkan ran Qwen3 8B at about 13.4 tokens per second, so on AMD the two land in the same range. On a single Lunar Lake chip, a Core Ultra 7 258V, one Gemma test ran at 24.9 tokens per second on the NPU against 36.0 on the Arc 140V. The NPU’s win is lower power draw and keeping the GPU free. If token speed is the goal, read our local AI mini PC guide instead, where memory bandwidth decides it.
In a homelab
Frigate can run object detection on an Intel NPU through OpenVINO, at about 6 to 40 ms per inference depending on the model, and its docs list a community AMD XDNA2 detector too. That’s a real alternative to the accelerators in our Hailo vs Coral comparison. Immich’s machine learning does not use Intel or AMD NPUs. Whisper speech-to-text runs on the AMD NPU through Lemonade and on Intel through OpenVINO, though Intel’s NPU path still has open bugs.
How to choose a mini PC with an NPU
- Read NPU TOPS, not platform TOPS. 40 is the Copilot+ line. AMD’s Ryzen AI chips are 50 to 60 (55 on the HX 470 in our top pick), Intel’s Lunar Lake and Panther Lake 40 to 50, and Snapdragon X2 Elite 80.
- Do the barebone math. A 32GB DDR5 SO-DIMM kit like the TeamGroup Elite 2x16GB was about $480 when we checked. Add that and an SSD before you compare a barebone with a built box. Our DDR5 RAM guide has more kits.
- Insist on dual channel. One SO-DIMM halves memory bandwidth, which hurts the iGPU and any local AI work. Built systems sometimes ship with a single stick.
- Soldered RAM is now a price advantage. Lunar Lake, big-GPU Panther Lake and Snapdragon boxes include their memory, so they dodge the shortage, but you cannot upgrade them later.
- For Linux, buy XDNA2 or a recent Intel chip. FastFlowLM and Lemonade need Ryzen AI 300 or newer. Intel works through OpenVINO. Skip Snapdragon until Qualcomm’s NPU support lands.
- Check Recall’s extra requirements. A biometric sign-in and Secured-core firmware decide whether Recall works, even on a box with the right chip.
Which mini PC with an NPU to buy
- One box for Windows AI features and future upgrades: Minisforum AI X1 Pro-470.
- The cheapest complete Copilot+ PC: GMKtec K13.
- An office fleet with a warranty: ASUS ExpertCenter PN54.
- The smallest footprint: MSI Cubi NUC AI+ 3MG.
- 10GbE and the strongest Intel graphics: Minisforum M2 Pro.
- The fastest NPU and every Copilot+ feature: ASUS Ascent QN10.
- macOS instead of Windows: Mac mini (M6).
If you mainly want local LLMs at a usable speed, an NPU should not be what you’re paying for. Buy for memory bandwidth and GPU instead, and treat the NPU as a bonus for background jobs. If the budget is closer to a Raspberry Pi than a mini PC, our edge LLM comparison covers the Jetson Orin Nano Super, the AI HAT+ 2 and the Hailo-10H.
