GMKtec EVO-X2 Review: A Mini PC That Runs 120B AI Models

The high-end mini PC market has changed shape in a matter of months. Where these machines once shipped with 32 GB of RAM and a modest integrated graphics chip, the GMKtec EVO-X2 arrives on AMD's Strix Halo platform — silicon originally intended for professional laptops and compact workstations — and pushes the boundaries of what a mini PC can do. Inside sits an AMD Ryzen AI Max+ 395 with 16 Zen 5 cores and 32 threads, paired with a Radeon 8060S iGPU carrying 40 RDNA 3.5 compute units, backed by 128 GB of unified LPDDR5X-8533 memory. It is that memory, and the local-AI work it unlocks, that makes this machine so unusual.

The EVO-X2 stays compact despite its power. The aluminium chassis, with bronze-gold flanks, feels dense and well finished, with discreet fan lighting and a front button to switch lighting and fan profiles. Connectivity is generous: front USB-A 3.2 Gen 2 and a USB4 port (40 Gbps, DisplayPort Alt Mode, Power Delivery), plus two more USB4, HDMI 2.1, DisplayPort 1.4 and 2.5 GbE at the rear — enough to drive up to four displays, though there is no OCuLink for an external GPU. The external power supply is a hefty 230 W brick, in keeping with a machine that can pull up to 140 W at peak, and a VESA mount is included.

Opening it up is a little fiddly, with screws hidden under glued rubber feet, but inside is an ambitious "Max3.0" cooling system: three copper heat pipes, two sizeable CPU fans and a third dedicated to the M.2 SSDs. The large, slow-spinning main fan moves plenty of air quietly. Two M.2 2280 PCIe 4.0 slots accept up to 16 TB of storage, though the memory is soldered — an unavoidable trait of Strix Halo, and hardly limiting at 128 GB.

The Ryzen AI Max+ 395's 16 cores and 32 threads make the difference in parallel work. In testing it scored 22,009 in Geekbench 6 multi-core, well ahead of the Intel Core Ultra X7 358H (15,665), and 7,042 in Cinebench 2026 multi — around 1.6 times the X7. Single-core is closer, with Intel's Panther Lake keeping a slight edge, so the AMD chip's advantage is firmly in compilation, rendering, encoding and containers rather than everyday snappiness. Thermals are healthy: under a sustained 32-thread load the clock held a flat plateau with no throttling, the sensor reading around 39 °C while the cooler absorbed the chip's 124 W.

The Radeon 8060S is the other headline. Its 40 compute units — against the 12 to 16 of most rival iGPUs — deliver performance comparable to a mobile RTX 4060. Games run well at 1440p: Cyberpunk 2077 averaged 74 fps on Ultra with FSR Quality, and with FSR3 frame generation the most demanding titles clear 120 fps. For an integrated GPU in a box this small, that is unprecedented.

The real vocation, though, is local AI. The 128 GB of unified memory sits on a 256-bit, eight-channel bus, and a BIOS "VRAM carveout" reserves part of it for the GPU — 64 GB by default, leaving 64 GB for the system. Sixty-four gigabytes of dedicated graphics memory on a mini PC is extraordinary, more than most professional graphics cards. The payoff is dramatic: every model tested ran entirely on the GPU. The 120-billion-parameter gpt-oss model ran at 35.7 tokens per second — a 120B model running wholly on a mini PC's integrated graphics, at a usable conversational speed, which normally demands a professional card costing several thousand pounds. Mixture-of-Experts models such as qwen3:30b hit 92 tokens per second, while dense models are slower but still usable (a 70B Llama managed 5.3). Notably the same chip in a rival machine ran roughly half as fast, purely because its BIOS defaulted to a 2 GB carveout that pushed work onto the CPU; on Strix Halo, the BIOS setting is everything.

Software choice matters too. Windows and Linux performed identically at the same backend, but on AMD's GPU the generic Vulkan backend proved 16 to 21 per cent faster than AMD's own ROCm for text generation, whereas ROCm won on processing long prompts — so Vulkan suits chat, ROCm suits document-heavy work.

The bundled 2 TB NVMe PCIe 4.0 SSD is quick where it counts, posting 595,000 read and 470,000 write 4K random IOPS, and a free second slot allows up to 16 TB. Networking runs to Wi-Fi 7 (a real 1.26 Gbit/s in testing) and 2.5 GbE. The trade-off for all this power is consumption — the CPU and GPU share a package envelope of about 130 W, roughly double a conventional mini PC — and noise, measured at 47 dB at one metre under full load, though it is a soft, even rush rather than a high-pitched whine.

At the time of writing the EVO-X2 starts at around £2,099.96 at Amazon UK for the entry configuration, with the tested 128 GB/2 TB model higher (about £3,666). Sold direct and via Amazon, it is squarely a specialist machine: superb for AI developers and researchers who want local inference of 70B or even 120B models without sending data to the cloud, for content creators wanting serious CPU and GPU power without a tower, and for enthusiasts who want 1080p/1440p gaming from a tiny box. It is wasted on pure office use, cannot match a real graphics card for 4K ray-traced gaming, and its memory is not upgradeable. The niggles are a high price, up to 130 W draw, the awkward teardown and a 12-month warranty. But for a compact local-AI workstation with excellent cooling, the EVO-X2 is one of the best options on the market.