Mini PCs have become one of the more important segments of the client computing market. They anchor fleets of clutter-free office desks, drive digital signage and kiosks, and increasingly sit at the edge running inference close to cameras and sensors where cloud round-trips are too slow or too expensive. What the category has lacked until now is a highly compelling Arm option, and the ASUS Ascent QN10 supplies one as the first mini PC built on Qualcomm’s Snapdragon X2 Elite, a machine ASUS bills as the world’s first mini PC with an 80 TOPS NPU.
The story of the QN10 starts with its silicon, because the X2E-88-100 pairs 18 third-generation Oryon cores at up to 4.7GHz with the Adreno X2-90 GPU and the 80 TOPS Hexagon NPU, all in a chassis that measures under 0.7 liters, displacing about as much desk space as a stack of granola bars. Despite the size, ASUS fits a vapor chamber, a dedicated fan for the PCIe Gen5 SSD slot, seven USB ports including three USB4, HDMI 2.1, 2.5GbE, and Wi-Fi 7, with support for four simultaneous 4K displays. For enterprise buyers, fTPM 2.0 and Qualcomm’s SPU with Microsoft Pluton support cover security, and the platform’s performance-per-watt is the pitch behind the 180W adapter: this is a Copilot+ desktop that sips power at idle and spins its fan to zero RPM.
ASUS positions the QN10 at AI developers, prosumers, and edge deployments: local inference for signage and kiosks, computer vision for industrial and retail settings, and always-on assistant workloads, backed by ready-to-run models through the Qualcomm AI Hub. The 16GB/512GB configuration sells for $1,349.99 at Best Buy at this writing; we tested the 32GB build. Benchmarking a Windows-on-Arm desktop still means navigating what actually runs natively, so our suite here is the Arm-compatible subset, and the comparison set reflects the category’s awkward moment: no other Snapdragon X2 desktop exists yet.
ASUS Ascent QN10 Specifications
| Normen | ASUS Ascent QN10 |
|---|---|
| Modell | ASUS Ascent QN10 |
| Prozessor | Qualcomm Snapdragon X2 Elite (X2E-88-100), 3rd Gen Oryon CPU, 18 cores / 18 threads, up to 4.7GHz |
| Grafiken | Qualcomm Adreno X2-90 (integrated) |
| NPU | Qualcomm Hexagon NPU, 80 TOPS (INT8) |
| Memory | 32GB LPDDR5x, 8533 MT/s, onboard (16GB and 32GB configs; 32GB is the ceiling) |
| Lagerung | 512GB SanDisk PC SN5100S as tested; 1x M.2 2280 PCIe Gen5 + 1x M.2 2280 PCIe Gen4 slots, up to 4TB total |
| Netzwerken | Wi-Fi 7, Bluetooth 6.0, Realtek 2.5GbE LAN |
| Ports (front) | 2x USB4 Type-C (DP1.4/PD, 40Gbps) 1x USB-A 3.2 1x USB-A 2.0 1x 3.5-mm-Audiobuchse |
| Ports (rear) | 1x USB4 Type-C (DP1.4/PD, 40Gbps) 2x USB-A 3.2 1x HDMI 2.1 FRL 1x RJ45 2.5GbE |
| Unterstützung anzeigen | Up to 4 displays (HDMI + 3x USB-C) |
| Sicherheit | fTPM 2.0, Qualcomm SPU with Microsoft Pluton support, Snapdragon Guardian |
| Kühlung: | CPU fan with vapor chamber, dedicated Gen5 SSD fan; max 53 dBA at full speed, 0 RPM at idle |
| Tuning | 180W DC adapter |
| Betriebssystem | Windows 11 Pro, Copilot+ PC |
| Abmessungen / Gewicht | 130 x 130 x 40mm (43.5mm with feet), under 0.7L / 620g |
| Langlebigkeit | MIL-STD 810H tested (shock, vibration, humidity, temperature, port stress) |
| Preis | $1,349.99 (16GB/512GB, Best Buy at this writing); the 32GB/512GB configuration as tested is listed at Newegg but out of stock with no price shown |
Where the QN10 Fits in ASUS’s Mini PC Lineup
ASUS took over Intel’s NUC business in 2023 and has kept the x86 line on a steady cadence since: the Meteor Lake NUC14Pro we reviewed, the current Arrow Lake NUC 15 Pro, and performance tiers above them that pair Arrow Lake with discrete RTX graphics for gaming and creator work. That NUC Pro line is the category default for a reason: barebones kits with socketed memory and storage, the full weight of x86 Windows compatibility, and a price floor that a self-configured build keeps under $850 today.
Ascent is the newer, AI-first branch of the family, and the QN10 is its second effort. The first was the Ascent GX10, which put NVIDIA’s Grace Blackwell GB10 in a desk-side box for CUDA developers working against datacenter toolchains. The QN10 takes the other lane: Qualcomm silicon, a Windows Copilot+ stack, and an NPU-first design aimed at inference that runs all day at low power rather than model development. Where the GX10 is a small workstation for building AI, the QN10 is closer to a mini appliance for running it.
The Arm against x86 in this lineup follows that split: the NUC Pro remains the pick when configurability, upgrade paths, or unconditional application compatibility decide the purchase; the QN10’s case is performance-per-watt, the 80 TOPS NPU, and sealed, deploy-and-forget roles like signage, kiosks, and edge inference, with enough CPU behind it, as the results below show, that the efficiency story no longer requires a performance apology.
Aufbau und Design
The QN10 is a 130mm square, 40mm tall, and 620 grams, displacing under 0.7 liters. The chassis is a plain silver metal body with an ASUS wordmark on the front, which suits the deployment scenarios this machine is built for. It’s built to disappear behind a monitor or into a kiosk enclosure, which is exactly where ASUS expects it to live.
The front panel carries the day-to-day connections: two of the three USB4 Type-C ports (DP1.4 and Power Delivery, 40Gbps), a USB-A 3.2 port alongside a USB-A 2.0 port, a 3.5mm audio jack, and the power button. Putting two full USB4 ports on the front is a small deployment kindness; docks, capture hardware, and external NVMe all connect without reaching around the back.
The rear holds the third USB4 Type-C, two more USB-A 3.2 ports, HDMI 2.1 with FRL, the RJ45 for the Realtek 2.5GbE controller, and the DC input for the external 180W adapter. Between HDMI and the three USB4 ports, the QN10 drives up to four displays, and with seven USB ports total, it gets through a full desk or signage installation without a hub.
The right side of the chassis is a ventilation inlet, feeding a cooling system that is genuinely overbuilt for the class: a CPU fan over a vapor chamber, plus a second, dedicated fan for the PCIe Gen5 SSD bay. ASUS rates the box at a maximum of 53 dBA at full fan speed, and at idle the fans stop entirely at 0 RPM.
Opening the chassis shows where the thermal budget goes and what a buyer can and cannot change, since the shell carries copper thermal pads that mate against the M.2 drives, one PCIe Gen5 slot (labeled G5, empty on our unit), and one Gen4 slot holding the SanDisk drive, expandable to 4TB total (though the user should be able to load higher capacity if they choose), while the mainboard packs the Snapdragon package under the vapor chamber with the 32GB of LPDDR5x soldered alongside. Storage is the only field-serviceable component; memory is fixed at purchase, and 32GB is the max.
With the fan lifted off, the vapor chamber spans the SoC and feeds a single heat pipe loop, a laptop-grade solution in a desktop where sustained load, not battery, is the constraint. The benchmarks that follow suggest it works, and anecdotally so do our ears: from idle through sustained heavy load, the QN10 stays whisper quiet, with the fans barely audible from three feet away. The only time they made themselves heard in our testing was under extreme system stress, such as a BIOS update.
ASUS Ascent QN10 Performance
Our review unit runs the Snapdragon X2 Elite X2E-88-100 with 32GB of LPDDR5x and a 512GB SSD on Windows 11 Pro, with all benchmarks run on the High performance power plan. Windows on Arm constrains the benchmark table: everything below ran natively on ARM64 where a native build exists, and the AI workloads note which inference engine and compute device each run used, because engine choice changes these numbers as much as silicon does. PCMark 10 is absent because its main productivity benchmark has no Arm build; only its storage and battery tests run natively, and neither tells this desktop’s story.
Comparables are straightforward but imperfect, because nothing else in the lab matches this machine’s architecture and class at once. The ASUS NUC 14 Pro (Core Ultra 7 165H) is the x86 mini PC yardstick because it is the one we have: ASUS has since moved the line on to the Arrow Lake NUC 15 Pro, which we have not tested, so the Meteor Lake NUC 14 is standing in for x86 a generation back. It acquits itself better than its age suggests in the results below, and pricing keeps the comparison interesting in a way the release calendar does not. Our NUC 14 configuration, the Core Ultra 7 165H with 16GB and 512GB, listed at $1,522.61 at Best Buy, has now aged out of new-condition retail, while the QN10’s equivalent 16GB/512GB build sells for $1,349.99, so configured-for-configured the Arm box enters below where this class of x86 mini PC actually sold. The counterpoint is the do-it-yourself floor: a current NUC 15 Pro barebones kit runs about $700 at this writing and lands under $850 with memory and storage added, a path the sealed, soldered QN10 cannot match. The NUC’s 16GB memory ceiling limits the largest y-cruncher runs, and a few tests that failed or were skipped on that platform are noted where they occur. The Lenovo ThinkCentre Neo 50q QC is the closest Arm relative we have tested, built on the previous-generation Snapdragon X; that unit has left the lab, so its comparisons are limited to the benchmarks we published at the time, and we re-ran several older tests on the QN10 specifically to line up against it. Newer tests like Geekbench 7 have no Lenovo column for that reason and are presented for context.
Geekbench 7
Geekbench 7 joins the suite alongside Geekbench 6 as comparison data builds. Its CPU scores are calibrated against a baseline of 2,500, set by the AMD Ryzen 7700, while GPU scores are calibrated against a baseline of 100,000, set by the NVIDIA GeForce RTX 4060. Higher scores are better, and double the score indicates double the performance. Because Geekbench 7 uses new workloads and new baselines, its scores are not comparable to Geekbench 6 results. It shipped after the Neo 50q left the lab, so the Lenovo column stays empty and these comparisons run against the freshly re-benched NUC 14 Pro. The QN10 ran the native AArch64 build.
| Geekbench 7 (höher ist besser) | ASUS Ascent QN10 | ASUS NUC 14 Pro (Core Ultra 7 165H) | Lenovo ThinkCentre Neo 50q QC |
|---|---|---|---|
| CPU Single-Core | 3,272 | 2,291 | N / A |
| CPU-Mehrkern | 24,049 | 13,911 | N / A |
| GPU OpenCL | 30,982 | 27,790 | N / A |
| GPU Vulkan | 40,042 | 25,088 | N / A |
The headline numbers: 3,272 single-core and 24,049 multi-core, 43% and 73% ahead of the NUC 14 Pro’s Core Ultra 7 165H on the same suite. For scale against our laptop stack, that multi-core figure also clears every Panther Lake system we have tested this year, machines that cost three to four times as much. The Adreno X2-90 lands at 40,042 in Vulkan, 60% past the NUC’s Arc, and 30,982 in OpenCL, in the range of 12 Xe-core Arc B390 laptops rather than basic integrated graphics.
Geekbench 6
Geekbench 6 measures processor performance using a mix of common tasks, with separate scores for single-core and multi-core workloads, plus GPU compute scores through OpenCL and Vulkan. Higher scores are better. It is a generation older than Geekbench 7, which is exactly why it is here: it is the suite we published for both comparison systems, and we re-ran it on the QN10 through the native Arm build to line up against them.
| Geekbench 6 (höher ist besser) | ASUS Ascent QN10 | ASUS NUC 14 Pro (Core Ultra 7 165H) | Lenovo ThinkCentre Neo 50q QC |
|---|---|---|---|
| CPU Single-Core | 3,667 | 2,517 | 2,148 |
| CPU-Mehrkern | 19,884 | 12,477 | 8,565 |
| GPU OpenCL | 39,761 | 35,449 | 9,634 |
| GPU Vulkan | 43,942 | 36,089 | N / A |
Against the previous-generation Snapdragon X in the Neo 50q, the QN10 posts 3,667 single-core against 2,148 and 19,884 multi-core against 8,565, a 2.3x multi-core jump in one Arm generation. The x86 comparison is just as lopsided: the NUC 14 Pro’s Core Ultra 7 165H manages 12,477 multi-core on its re-run, 63% of the QN10’s score. GPU compute tells the same story, with the Adreno at 39,761 in OpenCL against the Neo 50q’s 9,634 and ahead of the NUC’s Arc at 35,449.
Cinebench 2026
Cinebench 2026 is the current release in the Cinebench line and the only version we report. It tests CPU and GPU performance using Maxon’s Redshift render engine, built on the latest Cinema 4D 2026 code, and is designed to show whether a machine is stable under high CPU load, whether its cooling can sustain longer render tasks, and how it handles demanding real-world 3D work. Because code and compiler changes accelerated scene rendering, Cinebench 2026 scores use an adjusted range and should not be compared to scores from previous Cinebench versions. The benchmark ships a native ARM64 build, which the QN10 ran; its GPU test does not support the Adreno or the NUC’s Intel integrated graphics, so only the CPU results appear here.
| Cinebench 2026 (höher ist besser) | ASUS Ascent QN10 | ASUS NUC 14 Pro (Core Ultra 7 165H) | Lenovo ThinkCentre Neo 50q QC |
|---|---|---|---|
| CPU-Single-Thread | 639 | 444 | N / A |
| CPU mit mehreren Threads | 6,476 | 3,684 | N / A |
| MP-Verhältnis | 10.13x | 8.29x | N / A |
At 639 single-thread, the QN10 posts the highest Cinebench 2026 single-thread score we have recorded on any system in this class; Maxon’s own reference chart places it between the Apple M4 Max and Intel’s desktop Core Ultra 9 285K. Multi-thread lands at 6,476 with a 10.13x MP ratio against the NUC 14 Pro’s 3,684 at 8.29x, a 76% lead, and the 639-to-444 single-thread gap holds the same shape, which makes both results remarkable for a 0.7-liter box.
3DMark CPU-Profil
The 3DMark CPU Profile benchmark measures CPU performance at fixed thread counts, from a single thread up to the maximum available, showing how performance scales as more cores are engaged. Higher scores are better. It is one of the tests we published for the Neo 50q, so we re-ran it on the QN10 to keep that comparison live.
| 3DMark-CPU-Profil (höher ist besser) | ASUS Ascent QN10 | ASUS NUC 14 Pro (Core Ultra 7 165H) | Lenovo ThinkCentre Neo 50q QC |
|---|---|---|---|
| Maximale Threads | 9,544 | 7,680 | 3,370 |
| 16-Threads | 9,118 | 6,899 | 3,360 |
| 8-Threads | 5,327 | 5,617 | 3,497 |
| 4-Threads | 3,209 | 3,582 | 2,152 |
| 2-Threads | 1,702 | 1,914 | 1,422 |
| 1 Thema | 857 | 1,002 | 712 |
The generational gap is stark at every point on the curve: 9,544 max threads against the Neo 50q’s 3,370, and 857 single-thread against 712. The 8-thread crossover is worth noting: 5,327 against 3,497, since most desktop workloads live in that range. The curve also flattens noticeably between 16 threads and max on the QN10, the expected shape for an 18-core part without SMT. Against the NUC 14, the picture is more nuanced than the Geekbench results suggest: the QN10 wins the top of the curve, 9,544 to 7,680, but the 165H takes the single-thread point, 1,002 to 857, and holds a slight edge through 8 threads. This workload rewards burst clocks in ways the Oryon design does not chase, which is worth remembering for lightly threaded desktop work.
3DMark Graphics
The 3DMark graphics suite covers a spread of rendering workloads: Solar Bay and Wild Life are cross-platform tests built for integrated and mobile-class graphics, Steel Nomad and Time Spy are heavier DirectX 12 rasterization tests, Fire Strike is the DirectX 11 benchmark, and Port Royal and Speed Way exercise DirectX Raytracing. Higher scores are better throughout. Solar Bay, Steel Nomad, and Wild Life overlap with our published Neo 50q data; the rest establish the QN10’s baseline against the re-benched NUC 14 Pro, including the first DXR ray-tracing runs we have completed on a Windows-on-Arm box.
| 3DMark Graphics (higher is better) | ASUS Ascent QN10 | ASUS NUC 14 Pro (Core Ultra 7 165H) | Lenovo ThinkCentre Neo 50q QC |
|---|---|---|---|
| Solarbucht | 21,561 | 11,630 | 5,971 |
| Stahlnomad | 1,042 | 611 | 235 |
| Tierwelt | 19,996 | N / A | 11,224 |
| Zeit Spion | 4,014 | 3,643 | N / A |
| Feuer-Schlag | 9,949 | 6,591 | N / A |
| Port Royal | 1,520 | 1,429 | N / A |
| Geschwindigkeitsweg | 324 | 335 | N / A |
Solar Bay at 21,561 is 3.6x the Neo 50q and 1.9x the NUC 14, and Steel Nomad at 1,042 is 4.4x the previous Arm generation and 70% past the Arc. Time Spy at 4,014 and Fire Strike at 9,949 put the QN10 in entry-discrete territory rather than the bottom of the iGPU pile. Ray tracing is the one place the Arc claws back: Port Royal favors the QN10 narrowly at 1,520 to 1,429, while Speed Way goes to the NUC 335 to 324, effectively a tie. Nobody buys either box for ray tracing, but the capability checkbox on Arm is real. The NUC 14 column has no Wild Life entry because that test was not part of its re-bench pass.
3DMark-Speicher
3DMark Storage measures how an SSD performs during gaming-related tasks such as loading games, installing software, saving progress, and moving game files, and it runs natively on Arm. Higher is better. The score reflects the drive each system shipped with, so this is a comparison of the vendors’ storage choices rather than the platforms.
| 3DMark-Speicher (höher ist besser) | ASUS Ascent QN10 | ASUS NUC 14 Pro (Core Ultra 7 165H) | Lenovo ThinkCentre Neo 50q QC |
|---|---|---|---|
| Storage Score | 2,393 | 1,887 | 2,871 |
At 2,393, the QN10’s 512GB SanDisk PC SN5100S lands below the Neo 50q’s 2,871 but ahead of the NUC 14’s 1,887. The SN5100S is a Gen4-class client drive and sits in the Gen4 slot; the Gen5 slot, labeled G5 on the board, ships empty and is where a faster drive would go, since the as-shipped 512GB unit is serviceable.
7-Zip-Komprimierung
The built-in 7-Zip benchmark measures how quickly the processor can compress and decompress data using multiple threads, run with a 128MB dictionary across ten passes. Decompression tends to scale with thread count while compression leans on memory latency, so the two halves often tell different stories. Higher GIPS scores are better. 7-Zip 26.03 ships a native ARM64 build, which the QN10 used.
| 7-Zip (GIPS, higher is better) | ASUS Ascent QN10 | ASUS NUC 14 Pro (Core Ultra 7 165H) | Lenovo ThinkCentre Neo 50q QC |
|---|---|---|---|
| Komprimieren | 96.729 | N / A | 53.096 |
| Dekomprimieren | 120.205 | N / A | 46.998 |
| Gesamtbewertung | 108.467 | N / A | 50.047 |
The QN10’s 108.467 GIPS total more than doubles the Neo 50q’s 50.047, with decompression at 120.205 GIPS the standout half of the run. For calibration, that total edges out every laptop in our recent Panther Lake stack; this is real multi-core throughput, not a scaled-down mobile result. The NUC 14 column is empty because 7-Zip’s benchmark did not produce a result on that system in our original review, and it was not part of the re-run.
Y-Cruncher
y-cruncher measures how quickly the processor can calculate large numbers of digits of Pi, placing a heavy load on the CPU and memory subsystem, while the BBP runs extract hexadecimal digits of Pi in a workload that is almost purely CPU-bound. Results are in seconds, so lower times are better. The benchmark runs natively on Arm. The QN10’s 5-billion-digit run was skipped due to available memory, the expected ceiling for a 32GB machine, and the 16GB NUC 14 could not attempt anything past 1 billion digits in the standard runs.
| y-cruncher (Sekunden, je niedriger, desto besser) | ASUS Ascent QN10 | ASUS NUC 14 Pro (Core Ultra 7 165H) | Lenovo ThinkCentre Neo 50q QC |
|---|---|---|---|
| Pi 1B | 71.531 | 36.561 | 186.905 |
| Pi 2.5B | 218.332 | N / A | 497.704 |
| Pi BBP 1B | 5.844 | 2.306 | N / A |
| Pi BBP 10B | 67.701 | 25.144 | N / A |
| Pi BBP 100B | 778.722 | 296.970 | N / A |
The QN10 completes 1 billion digits in 71.5 seconds, 2.6x faster than the Neo 50q’s 186.9, though the NUC 14 Pro at 36.6 seconds shows Intel’s memory subsystem still wins this bandwidth-bound workload by nearly 2x. The BBP runs, which are pure CPU, tell the opposite story at the large end: the NUC finishes 100B hex digits in 297.0 seconds against the QN10’s 778.7. The 16GB NUC could not attempt anything past 1B in the standard runs, so the 2.5B comparison belongs to the Lenovo, where the QN10’s 218.3 seconds against 497.7 holds the 2.3x generational margin.
Mixer
The Blender benchmark measures rendering performance using three different 3D scenes: Monster, Junkshop, and Classroom. Results are reported in samples per minute, so higher scores are better. Scores are not comparable across Blender versions, so we report only the current Blender 5.2 results. Blender 5.2 ships a native ARM64 CPU path, which the QN10 ran, but the Adreno GPU is not a supported render device, so unlike our laptop reviews , this table is CPU-only for both systems.
| Blender 5.2 CPU (samples/min, higher is better) | ASUS Ascent QN10 | ASUS NUC 14 Pro (Core Ultra 7 165H) | Lenovo ThinkCentre Neo 50q QC |
|---|---|---|---|
| Monster | 188.04 | 109.16 | N / A |
| Trödelladen | 133.21 | 81.08 | N / A |
| Klassenzimmer | 93.93 | 57.41 | N / A |
The 18 Oryon cores render Monster at 188 samples per minute against the NUC 14’s 109, a 72% lead that holds across all three scenes, and the figure beats every x86 thin-and-light CPU result in our recent laptop stack. The missing GPU path matters, though: renderers that lean on the GPU will leave performance on the table here, and that is a software-ecosystem gap, not a hardware one.
UL Procyon KI
UL Procyon’s AI suite covers three workloads. AI Image Generation measures Stable Diffusion inference from low-power NPUs to high-end GPUs; we report the Stable Diffusion 1.5 INT8 test built for low-power accelerators. AI Text Generation runs local LLMs and scores generation performance, reporting time to first token and tokens per second; the Arm build covers Phi and Llama3. AI Computer Vision measures inference across image classification, object detection, segmentation, and super-resolution models, and the newer Computer Vision 2 suite runs through each vendor’s native path. Higher scores are better throughout. The engine labels matter more here than anywhere else in the review: on the QN10, image generation and text generation ran on the Hexagon NPU through Qualcomm’s QNN stack, computer vision ran through SNPE on the NPU and again through WinML, and the NUC 14 ran its best available paths, OpenVINO on the Arc iGPU and WinML. None of these numbers are comparable to the OpenVINO and ONNX results in our x86 laptop reviews, because Procyon scores only compare within the same engine and device.
| Procyon AI (higher is better) | ASUS Ascent QN10 | ASUS NUC 14 Pro (Core Ultra 7 165H) | Lenovo ThinkCentre Neo 50q QC |
|---|---|---|---|
| Image Generation, SD 1.5 INT8 | 5,457 (QNN, NPU) | 1,004 (OpenVINO, iGPU) | N / A |
| Text Generation Phi | 1,394 (QNN, NPU) | 528 (OpenVINO, iGPU) | N / A |
| Text Generation Llama3 | 1,371 (QNN, NPU) | 434 (OpenVINO, iGPU) | N / A |
| Computer Vision | 4,254 (SNPE, NPU) | 135 (WinML, GPU) | N / A |
| Computer Vision 2 (WinML) | 2,161 | N / A | N / A |
| Computer Vision 2 (SNPE) | 1,897 | N / A | N / A |
Taken on its own terms, the 80 TOPS Hexagon delivers: 5,457 in SD 1.5 INT8 image generation is roughly 1.9x what we have measured from 50 TOPS-class x86 NPUs on their own best engines, and text generation sustains 35.7 tokens per second on Phi and 25.2 on Llama3 with sub-0.6-second first tokens, entirely on the NPU with the CPU untouched. The hybrid HTP+CPU text generation run scored slightly lower than the pure NPU path, which suggests the NPU is not the bottleneck. Computer Vision 2 at 2,161 through WinML against 1,897 through SNPE is a rare case of the generic path beating the vendor path. The NUC 14 comparison underlines the caveat about engines: its best paths, OpenVINO on the Arc iGPU, land at 1,004 in image generation and 528 on Phi, and its NPU text generation run failed outright, so the QN10’s five-fold image generation lead mixes silicon and software advantages that cannot be fully separated.
Fazit
The ASUS Ascent QN10 is the first Arm mini PC we have tested that outruns its x86 class instead of trading performance for efficiency. Against the NUC 14 Pro, the Snapdragon X2 Elite leads by 43% to 76% across the CPU suites and roughly 2x in mainstream graphics, and it does that from a 0.7-liter box that stays whisper quiet from idle through sustained load. Against the previous Snapdragon generation in the Neo 50q, the jump is 2.3x to 4.4x depending on the test. The 80 TOPS NPU is not just a spec-sheet number either: 5,457 in SD 1.5 image generation and 35.7 tokens per second on Phi, all on the Hexagon with the CPU idle, is the kind of local inference edge-AI benefits from.
Our reservations start with memory: the 32GB of LPDDR5x is soldered, and 32GB is the ceiling, on a machine whose 16GB configuration already sells for $1,349.99 and whose 32GB build is listed, but not currently in stock, and fixed memory is a hard sell at this price because it turns the purchase into a decision the buyer can’t change in the future. A NUC 15 Pro barebones kit at about $700 takes whatever memory and storage a buyer wants to pair it with, and the x86 box keeps a few technical wins too. Storage is the milder complaint: the 512GB SanDisk SN5100S is a Gen4 drive that trails even the Neo 50q, though the empty Gen5 slot leaves the fix up to the buyer if more storage performance is needed. Windows on Arm compatibility remains a per-application question that any fleet buyer should understand before ordering.
The QN10 is still a great right machine for signage, kiosks, edge inference, always-on desk-side assistants, and developers building on the Qualcomm AI Hub, anywhere a quiet, efficient appliance with a capable NPU and 32GB is enough. ASUS has built a convincing first entry for the Ascent line’s Arm lane; it’s great to see Snapdragon looking much more viable in this round, and we’re enthusiastic to see more.
Amazon