Two Years Later, the Real Mini PC Arrives: Asus Ascent Q10 and Snapdragon X2 Elite Tested

Two years ago Snapdragon X Elite was shown and then pulled; only a few hundred kits shipped and buyers were refunded. That gap turned into a long wait for a real Windows on ARM desktop. In the meantime Project Voltera and its successor Windows Dev Kit 2023 already enabled native ARM64 for Visual Studio and .NET, yet they trailed on speed. Against the first X Elite laptop, Voltera scored 14 vs 26 on Speedometer and took about 90 seconds on the Python Mandelbrot, roughly half the pace. The delay pushed Qualcomm's volume ramp from 2025 to 2026.

Asus Ascent Q10: the first real X2 Elite desktop

The Asus Ascent Q10 (also press-named QN10) fills that gap as the first production box with Snapdragon X2 Elite . The review unit has 32 GB LPDDR5X and 512 GB storage , with three USB4 , 2.5 GbE Ethernet and Wi-Fi 7 on the back. Power comes from a bulky 180 W brick; the reviewer rightly flags it in 2026 and wonders whether the chip will ever need the full wattage. The box stays quiet, gets warm under load but not loud. Wall power is about 109 W during the build and 120-125 W for the Python run.

Anatomy of the X2 Elite 88-100

Inside is Snapdragon X2 Elite X2E-88-100 : third-gen Oryon on 3 nm , 18 cores (12 prime up to 4.7 GHz, 6 performance up to 3.4 GHz) and 53 MB cache . All cores are big cores, no little efficiency cores. The controller is LPDDR5X-9523, 128-bit, 152 GB/s ; the Extreme variant reaches 228 GB/s . For reference M4 120 GB/s , M4 Pro 273 GB/s . Graphics is Adreno X2-90 at 1.7 GHz , claimed 2.3x perf per watt over gen one; the Hexagon NPU at 80 TOPS (INT8) up from 45. The design scales the same architecture from laptop to desktop.

For developers the key is software. Visual Studio , .NET , Node.js and Python now ship as native ARM64 builds. Install the wrong binary and Prism translation kicks in, much like Rosetta on Apple, translating x64 on the fly with a clear slowdown. The video shows this trap twice: first Node, then Python in x64 form halving speed. Switching to the right ARM64 build doubles or triples the score. On Windows on ARM this choice is the most expensive mistake.

Synthetic scores: from Geekbench to browser fluidity

On synthetics Geekbench is clear: Q10 is 26 percent ahead single-core and 27 percent multi-core versus the cancelled X Elite kit, beating AMD HX470 and Intel Panther Lake boxes by 20-30 percent both ways, sitting within a hair of M4 Pro Mac Mini and edging M5 on MacBook Air (3597 single / 17550 multi) . The same day Speedometer 2/3 surprises: Q10 48.6 , M4 47.8 , M4 Pro 45 , HX470 37 marks the first time Apple trails on single-core fluidity; the score comes from the Chrome run, so it reflects everyday browser feel and with the right stack X2 Elite matches or passes the M4 family.

The companion web-tooling bench first confused. As a Node CLI running TypeScript/Babel/Prettier/Lebab, it gave 8.15 runs per second on x64 Node and 11.8 on ARM64 Node ; SER10 28, old X Elite kit 22. The same suite as Chrome HTML then hit 63 on TypeScript and 49.89 geometric mean , leading the pack. The delta is the V8 wrapper difference between Node and Chrome plus architecture match. The lesson: if you run Node, ARM64 is mandatory; otherwise even the fastest chip runs at a third of its potential.

Real developer loads: builds and Python

For compiled code a custom .NET build with 100,000 namespaces and real code in each tries to keep the compiler honest. It finishes in 77 seconds , quiet and barely warm, with the case only mildly hot. M4 Pro 66 seconds takes first, M4 106 , old X Elite kit 86. So Q10 is second on multi-core builds while staying quiet on the desk for hours. Power is around 109 W , so the 180 W brick is not saturated.

The interpreted Mandelbrot Python bench pushes every core to 100 percent. First with x64 Python: 50.43 seconds , fans just audible yet still quiet. With ARM64 Python: 21.76 seconds it wins: M4 Pro 23.2 , M4 31.4 , old kit and SER10 28.9. Wall power is 120-125 W . Together the two tests prove: with the right builds X2 Elite trails M4 Pro modestly on multi-core builds yet beats it on heavy single-work Python; with the wrong binary the same silicon pays double.

AI: NPU attempt, GPU limit and CPU surprise

The AI section tempers expectations. The 80 TOPS Hexagon NPU try with Llama.cpp + Hexagon backend stalled at the driver and never ran. Unified memory is 32 GB, but via OpenCL the GPU can use only about 15 GB , the other 17 GB stays reserved, so large models fall back to CPU. Curiously CPU beats GPU on every model here: Llama 3B CPU 41 vs GPU 31 tokens per second , GPT-OSS 20B MoE 35 vs 33 . For scale: 5 tokens per second is reading speed , 40 is faster than you can skim , 100 finishes before you look up .

Across models the picture sharpens. Llama 3.2 3B: Q10 41, old kit 36, M4 46, M4 Pro 101 (Apple runs MLX , its best case). Dense 27B models drop to 5-6 tokens per second at reading speed, waking every parameter each token. Mixture-of-experts (MoE) restores pace: GPT-OSS 20B 35 , Qwen 3 35B 27 tokens per second , and the 22 GB Qwen fits the 32 GB Q10 while it does not fit a 16 GB Mini. A realistic 16,000 token context waits 53 seconds to first token ; with a second user on the same box everyone falls to about 7 tokens per second , so today this is a one-person desk machine.

The closing verdict is mature. For Windows developers the Mac Mini killer is yes : it keeps the two-year promise and competes at the top for builds, browser and Python. For AI it is no : at a higher price the 599-dollar base M4 with MLX is faster on small models, bandwidth lifts M4 Pro on large ones, and Q10 costs more. Check ARM64 compatibility first. The M6 wave will reshuffle the chart, but today the Q10 is a quiet, efficient pick for teams that want power on a Windows desk.