GMKtec · Mini PCs

GMKtec EVO-X2 review: a desktop-sized local AI machine for the price of a gaming laptop

Ryzen AI Max+ 395 with up to 128 GB of unified memory runs 70B language models at home. Quiet at idle, ordinary value as a gaming PC.

8.5/10

Excellent

Best for Running large local AI models on your desk

List price$2,199.99

We may earn a commission when you buy through links on this page. It never changes our scores. Learn more

Preview review. Compiled from manufacturer specifications and published information. We will update the scores after hands-on testing.

Our verdict

The EVO-X2 is one of the cheapest ways to run big language models locally: with 128 GB of unified memory the GPU can use more than 50 GB, enough for a 70B model. ServeTheHome measured idle power of 8–14 W and 41–43 dBA under AI load. As a plain desktop or gaming PC it is expensive, and serviceability could be better. Buy it for AI, not for Steam.

What we like

  • Up to 128 GB unified LPDDR5X: the GPU can address more than 50 GB
  • Ran Llama 3.3 70B fully on the GPU at 3.7–3.8 tokens/s; Qwen3-32B at about 10.5 tokens/s
  • Low idle power (8–14 W) and quiet at idle (37–38 dBA)
  • Two USB4 ports, HDMI 2.1, DisplayPort, Wi-Fi 7, SD reader
  • Gaming between a desktop RTX 4060 and a laptop RTX 4070

What could be better

  • Expensive as a general desktop or gaming PC
  • Only one 2.5GbE port, no 10GbE
  • Screws hidden under the feet; Wi-Fi card hard to service
  • Gets audible under load: 43–44 dBA while gaming

Score breakdown

Performance
9.3
Build, cooling and noise
7.9
Ports and upgradability
8.1
Value
8.0

Overall = 35% Performance + 25% Build, cooling and noise + 20% Ports and upgradability + 20% Value. How we test

What is it?

The EVO-X2 is built around AMD’s Ryzen AI Max+ 395, nicknamed “Strix Halo”: 16 Zen 5 cores, a 40-CU Radeon 8060S graphics chip and up to 128 GB of fast LPDDR5X memory shared between CPU and GPU. That shared memory is the whole point. A normal graphics card has 8–24 GB of VRAM; here the GPU can use more than 50 GB, which is what large language models need.

GMKtec sells the 64 GB / 1 TB model for $2,199.99 and the 128 GB / 1 TB model for $3,499.99 at the time of writing.

Who it is for

Developers and tinkerers who want to run models like Llama 70B or Qwen 32B privately at home, without paying for cloud GPUs or a $10,000 workstation. If that is not you, keep reading the K8 Plus review instead.

Ports and build

ServeTheHome lists, at the front, a USB4 port, two USB 3.2 Gen 2 Type-A ports, an SD card reader and a headset jack, plus a performance-mode button. At the back are a second USB4 port, HDMI 2.1, DisplayPort, USB-A, USB 2.0 and a single 2.5GbE port. Wi-Fi 7 is built in.

Inside there are two M.2 slots, two blower fans and an RGB fan. The reviewer was not a fan of the layout: screws sit under the rubber feet, the Wi-Fi card is buried under an M.2 slot, and the top section is hard to service.

Local AI results

This is where the EVO-X2 earns its price. ServeTheHome’s 128 GB unit ran:

Model Setup Speed
Llama 3.3 70B Q6_K CPU only about 2 tokens/s
Llama 3.3 70B Q6_K Fully on GPU (over 50 GB) 3.7–3.8 tokens/s
Qwen3-32B GPU about 10.5 tokens/s

GMKtec’s own page claims much higher numbers for mixture-of-experts models (55 tokens/s for Qwen3 30B, 19 tokens/s for GPT-OSS 120B). Those are vendor figures with no test method, but they are plausible because MoE models only activate part of their weights. The practical takeaway: 30B-class models feel interactive, dense 70B models are slow but usable.

Gaming

ServeTheHome placed gaming performance between a desktop RTX 4060 and a laptop RTX 4070, with Counter-Strike 2 around 120 fps at 4K. That is impressive for integrated graphics, but the reviewer was clear: you would not buy this as a gaming PC at this price.

Power and noise

State Power Noise
Idle 8–14 W 37–38 dBA (34 dBA room)
Llama 70B inference 147–160 W 41–43 dBA
Gaming 170–180 W 43–44 dBA

Quiet at idle, noticeable under load. The reviewer wished it were a little quieter.

EVO-X2 vs the rivals

  • vs Framework Desktop: the same Ryzen AI Max+ 395 chip in a more repairable case. Compare current prices.
  • vs Apple Mac Studio: ServeTheHome compares it with a 512 GB M3 Ultra Mac Studio at around $10,000: more capacity, at a much higher price.
  • vs NVIDIA GB10 systems: 128 GB machines quoted from about $3,000, with NVIDIA’s CUDA software stack.
  • vs GMKtec K8 Plus ($919.99): for office work, light gaming and 3D-printing slicing, the K8 Plus does the job for less than half the price. See EVO-X2 vs K8 Plus.

Price and value

As ServeTheHome put it, about $2,000 feels like a lot for a basic desktop, but from an AI perspective it is a really strong system. We agree. Buy the 128 GB model if AI is the reason you are here; the 64 GB model limits you to smaller models.

Final thoughts

The EVO-X2 makes local 70B-class AI affordable and quiet enough for a desk. It is not perfect to work on, and it is not good value for gaming, but for its real job there is very little competition at the price.

Sources

Full specifications

CPUAMD Ryzen AI Max+ 395, 16 cores / 32 threads
GPURadeon 8060S, 40 CUs (RDNA 3.5)
Memory64 GB or 128 GB LPDDR5X (soldered, unified)
Storage1–2 TB NVMe, two M.2 slots
Front portsUSB4, 2× USB-A 3.2 Gen 2, SD card, audio
Rear portsUSB4, USB-A 3.2 Gen 2, USB 2.0, HDMI 2.1, DisplayPort, 2.5GbE
WirelessWi-Fi 7, Bluetooth
Power adapter230 W
Price64 GB/1 TB $2,199.99 · 128 GB/1 TB $3,499.99 · 128 GB/2 TB $3,649.99
Released2025

Ready to buy?

GMKtec EVO-X2

List price $2,199.99. Prices change often, so check the current deal.

Check price at GMKtec

Keep reading

More mini pcs

See all