Running large language models locally without cloud GPU rentals has typically demanded hardware costing tens of thousands of dollars. GMKtec's Evo-X5 Pro, first shown at IFA 2026 and now on sale, packs the AMD Ryzen AI Max+ Pro 495 and 192GB of LPDDR5X-8533 memory into a compact, rackable chassis. Up to 160GB of that unified memory pool can be allocated as VRAM, putting 300-billion-parameter models within reach on a single desktop machine. The Radeon 8065S integrated GPU gets roughly 273GB/s of memory bandwidth to work with.
The Gorgon Halo APU at the core of the Evo-X5 Pro brings 16 Zen 5 cores, 32 threads, boost clocks up to 5.2GHz, and 40 RDNA 3.5 compute units with hardware ray tracing. AMD's XDNA 2 NPU adds up to 55 TOPS for dedicated AI acceleration on top of the GPU and CPU. Connectivity is generous for a mini PC, with two USB4 v2 ports, two standard USB4, dual 10GbE Ethernet, HDMI 2.1, WiFi 7, and Bluetooth 5.4. Three M.2 slots support up to 24TB of total storage, and a triple-fan cooling system manages thermals. GMKtec also includes AMD DASH for remote management, which pairs well with the 2U rackable form factor for clustered deployments.
ROCm 7.14 added official support for the full Gorgon Halo family under Linux, including the Ryzen AI Max+ Pro 495 and its Radeon 8065S graphics. That means open-source inference tools like llama.cpp, Ollama, and vLLM can tap the full 160GB VRAM pool through AMD's compute stack, making this a practical option for self-hosted inference without relying on Windows or proprietary toolchains.
The Evo-X5 Pro ships in two configurations, 2TB of storage for $6,600 (€6,070) and 4TB for $6,900 (€6,350). That slightly undercuts the competing Minisforum MS-S1 Max-P495 while offering a comparable spec sheet. Both configurations are available now from GMKtec's website.
