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

The high-end mini-PC has changed beyond recognition in a matter of months. Where these one-litre boxes recently shipped with 32GB of RAM and a modest integrated graphics part, a new wave built on AMD's Strix Halo platform — originally meant for professional laptops and compact workstations — is redrawing the boundaries. The GMKtec EVO-X2 is one of them, and in extended testing it proved to be less a mini-PC than a local artificial-intelligence workstation that happens to fit in a shoebox. This review is based on that published laboratory test, not our own hands-on trial.

At the heart sits an AMD Ryzen AI Max+ 395: 16 Zen 5 cores and 32 threads up to 5.1GHz, a Radeon 8060S integrated graphics part with 40 RDNA 3.5 compute units, and an XDNA 2 neural engine — 126 TOPS of AI throughput all told. It is paired with 128GB of unified LPDDR5X-8533 on a 256-bit, eight-channel bus, soldered to the board, of which up to 64GB can be reserved as dedicated graphics memory in the BIOS. Storage is a fast 2TB PCIe 4.0 NVMe drive with a second M.2 slot free for up to 16TB total, and connectivity is generous: 2.5GbE, Wi-Fi 7, Bluetooth 5.4, twin USB4, HDMI 2.1, DisplayPort and up to four displays. A 45–120W power envelope is handled by three heat pipes and a twin-fan cooler.

On heavily threaded work the EVO-X2 is in a different class to Intel's latest small-form-factor processors. The test recorded 22,009 in Geekbench 6 multi-core against 15,665 for a Core Ultra X7 machine, and 7,042 in Cinebench 2026 multi — roughly 1.6 times the Intel part. For compilation, 3D rendering, heavy transcoding or containers, it takes a commanding lead. Single-thread is closer: Intel's Panther Lake keeps a slight edge that shows in desktop snappiness and some games. Thermals are healthy — under a sustained 32-thread load the clocks held a flat plateau with no throttling and a sensor reading around 39°C, the cooler shrugging off the processor's 124 watts.

The Radeon 8060S is the other headline. Its 40 compute units — where rival integrated graphics stop at 12 or 16 — deliver performance in the region of a mobile RTX 4060. Testers logged 11,289 in 3DMark Time Spy, Cyberpunk 2077 at 1440p Ultra averaging 74fps with FSR Quality, and Spider-Man 2 and Doom: The Dark Ages both clearing 60fps at 1440p; with FSR 3 frame generation the most demanding titles pass 120fps. For an integrated GPU in a mini-PC, that is unprecedented. Anyone building a full desktop instead should see our best gaming PCs guide and best desktop CPUs guide.

What sets the EVO-X2 apart is that unified memory. Because processor and graphics share one pool, the BIOS can carve out 64GB as graphics memory — more than most professional graphics cards — so large language models run entirely on the GPU. With that configuration every model tested ran fully on the GPU: a 3-billion-parameter Llama at 98 tokens per second, a 30-billion Qwen mixture-of-experts model at 92, and, most strikingly, a 120-billion-parameter model at a conversational 35.7 tokens per second. Running a 120-billion-parameter model on a mini-PC's integrated graphics is simply out of reach for any equivalent machine without a multi-thousand-pound professional card. Dense models are slower — a 70-billion Llama managed 5.3 tokens per second — but remain usable for considered writing or long code generation. Across the board the EVO-X2 roughly doubled a rival mini-PC running the same APU, thanks to sharper integration, cooling and BIOS tuning.

Testers ran the same suite under Windows 11 and Valve's Linux-based SteamOS. At a matched software backend the two operating systems performed identically — the reputation that Linux better exploits AMD graphics did not hold up. The bigger lever was the inference backend: AMD's own ROCm trailed the generic Vulkan by 16–21% on text generation, yet flipped to a commanding lead on prompt processing (up to 158% faster). The practical takeaway is clean — stay on Vulkan for chat and generation, switch to ROCm for heavy document processing and retrieval-augmented work.

The GMKtec EVO-X2 is a category-redefining machine. Its 128GB of unified memory and 64GB graphics carveout let it run models — up to 120 billion parameters — that no other mini-PC can touch, backed by workstation-grade multi-thread performance and a genuinely capable gaming iGPU built on AMD's Ryzen silicon. The caveats are honest ones: it is a specialist tool at a specialist price of around £2,100, single-thread performance sits a touch behind Intel, and the soldered memory cannot be upgraded later. For a compact, quiet local-AI workstation, though, nothing at this size comes close.