GMKtec EVO-X3 Review: 128GB Local AI Mini PC With OCuLink

The GMKtec EVO-X3 review from ServeTheHome, published on October 9, 2026, gives the first detailed numbers for one of the more interesting local AI mini PCs of the year. The machine pairs AMD's Ryzen AI Max+ 395 with 128GB of memory and adds something its rivals lack: an OCuLink port for external PCIe hardware. Reviewer Ryan Smith calls it "a solid entry in the AI workstation market."

• Processor: AMD Ryzen AI Max+ 395, 16 cores and 32 threads, with Radeon 8060S graphics.
• Memory: 128GB of soldered LPDDR5X-8000.
• Local AI result: 24.7 tokens per second on Phi 4 14B in MLPerf Client v1.6.1, with 1.46 seconds to first token.
• Price: the 128GB and 2TB configuration is going for $3,800, according to the review.

What Hardware Is Inside the GMKtec EVO-X3?

We covered the EVO-X3 early access launch in June. The review confirms the retail build. The chassis measures 353 x 186 x 41 mm and weighs 2.3kg, so it is a slim slab, not a cube. ServeTheHome notes the motherboard has unpopulated connectors for an LCD panel, a battery and a touchpad, which suggests the board began life as a laptop design.

Cooling is handled by a copper heatpipe assembly with two 6W blower fans that exhaust out the rear. A second baseplate covers the memory and power stages. Storage is a pair of M.2 2280 PCIe Gen4 x4 slots, with no built-in SSD cooling. Wireless comes from a replaceable MediaTek MT7925 card with Wi-Fi 7 and Bluetooth 5.4.

Ports, per the review:

• Rear: two HDMI 2.1, one 40Gbps USB-C, one 10Gbps USB-A, one 2.5GbE and one OCuLink (PCIe Gen4 x4).
• Front: one 5Gbps USB-A, one USB 2.0 and a 3.5mm combo jack.

How Fast Is It for Local AI?

The headline figure is MLPerf Client v1.6.1 running Phi 4 14B, which ServeTheHome describes as the most advanced model in the current client benchmark. The EVO-X3 averaged 24.7 tokens per second with a 1.46 second time to first token.

Power is the other half of the story. The firmware offers three modes: Silent at 54W, Balanced at 85W and Performance at 140W, and most testing used Performance. The review measured peaks of "around 160-180 Watts" under heavy sustained load from the 230W external supply, while LLM inference sat at "around 100 Watts or less." That is a modest draw for a machine with 128GB of memory available to models.

In Geekbench the EVO-X3 placed last among the Ryzen AI Max+ 395 systems ServeTheHome has tested, but the spread is small: within 3% in single-core and about 5% in multi-core. For practical purposes these machines perform alike, and the choice comes down to features and price.

Why Does the OCuLink Port Matter?

OCuLink carries a PCIe Gen4 x4 link outside the case. That opens the door to an external graphics card or other PCIe devices without the overhead of a USB4 tunnel, and ServeTheHome singles it out as the feature no other Ryzen AI Max mini PC offers. For local AI builders who want unified memory today and an add-in accelerator later, that is a useful upgrade path.

What Could Be Better?

The review lists a few areas for improvement. There is only one 40Gbps USB-C port, and it doubles as the sole DisplayPort output. Networking is 2.5GbE at a time when 10GbE is increasingly common on machines at this price. Smith also adds a fair caveat on fit: to get the most from it, "you really do need to be all-in on local AI inference."

That matches my read. If 128GB is the target, our guide to how much memory a local AI mini PC needs explains what that capacity buys, and the EVO-X3 is one of the less expensive ways to get it. More in our mini PC coverage.

Sources: ServeTheHome — GMKtec EVO-X3 Mini PC Review — October 9, 2026; ServeTheHome — performance and benchmarks — October 9, 2026.