GMKtec EVO-X5 Pro: 192GB at $6,799, or $35.41 per GB with 2TB

GMKtec sells its EVO-X5 Pro mini PC with 192 GB of unified memory for $6,799, and its US store listed the 2 TB model as in stock on Nov. 2. That works out to $35.41 per gigabyte of memory, against $54.30 for NVIDIA's 128 GB DGX Spark at the $6,950 NVIDIA's marketplace listed on Oct. 31 for the 4 TB model. Per-gigabyte figures here are Voltgrid arithmetic.

The box is built on AMD's Ryzen AI Max+ PRO 495, the chip AMD developed under the codename Gorgon Halo, and Level1Techs published a video about it on Oct. 31, "Gorgon Halo Arrives in GMKtec... Do you need 192 GB of LPDDR5?!?", with chapters comparing 192 GB with 128 GB and running benchmarks. This file quotes none of its figures until they appear with the model, quantization and engine attached. It sets GMKtec's and AMD's specifications against GMKtec's own cheaper boxes and against the machines Voltgrid has reviewed. Voltgrid did not test the machine.

What's New in the EVO-X5 Pro

The change that matters is memory. AMD's spec page gives the PRO 495 a 256-bit LPDDR5x bus at up to 8,533 MT/s and a ceiling of 192 GB. GMKtec fills it with eight 24 GB packages of onboard memory. Its earlier EVO-X2, on the Ryzen AI Max+ 395, tops out at 128 GB of LPDDR5X-8000.

More of that pool can go to the graphics processor. GMKtec and AMD say up to 160 GB can be allocated as video memory. Notebookcheck, reporting the launch on Sept. 28, set that against a 96 GB limit on the Ryzen AI Max+ 395. That is 64 GB more for model weights on the GPU, Voltgrid's arithmetic.

Speed barely moves. GMKtec's launch post rates the memory at up to 273 GB/s, which matches the bus width times the transfer rate on AMD's page. AMD lists 256 GB/s for the Ryzen AI Max+ 395 at 8,000 MT/s. The gain is about 6.6%, Voltgrid's arithmetic.

The rest of the chip is close to the 395. AMD lists 16 Zen 5 cores, a Radeon 8065S GPU with 40 compute units at up to 3,000 MHz, an NPU rated at up to 55 TOPS and up to 61 TFLOPS of FP16 compute. Notebookcheck describes the 8065S as clocked slightly higher than the 395's 8060S. The PRO label adds AMD DASH remote management, which GMKtec pairs with a discrete TPM 2.0 chip and a fingerprint reader.

The box around it is bigger than the EVO-X2. GMKtec lists 221 x 80 x 195 mm and about 2.5 kg, two 10 GbE ports where the EVO-X2 has one 2.5 GbE port, two 80 Gbps USB4 v2 ports, three M.2 slots for up to 24 TB, and a 330 W power adapter. Its Performance mode runs at 120 W sustained and 160 W peak, according to GMKtec. AMD's page gives the chip a configurable range of 45 W to 120 W, and GMKtec does not say how its 160 W peak relates to that range. Windows 11 Pro comes installed, and GMKtec says Linux and Ubuntu are supported.

What 192 GB Should Solve, and What It Costs

The promise is fit. GMKtec says the EVO-X5 Pro runs language models of up to 320 billion parameters fully offline, depending on the model, software and quantization. Its own lab table lists GLM-5.3-Flash in a 146.05 GB UD-IQ4_XS build decoding at 17.79 tok/s. That is a GMKtec lab figure, and its later posts claim higher speeds from its own software, which are vendor claims too. A file that size does not fit in the 96 GB a Ryzen AI Max+ 395 can hand its GPU.

Bandwidth sets the speed limit. Each new token reads the model's active weights from memory, so at 273 GB/s a dense model filling 160 GB could not exceed about 1.7 tokens per second, Voltgrid's arithmetic. The extra capacity pays off mainly on mixture-of-experts models, which read only a fraction of their weights per token. GMKtec's Oct. 30 post describes Qwen3.8-Flash-Next as a mixture-of-experts model of close to 180 billion parameters in total, and its table lists it at 93.48 GB and 45.08 tok/s decode, a GMKtec figure.

Price is where the box stands out, in both directions. Against the DGX Spark, the EVO-X5 Pro has 50% more memory at the same 273 GB/s and, with 2 TB against the Spark listing's 4 TB, costs $151 less. Like for like, GMKtec's 4 TB model costs $7,099, which is $149 more than the Spark but $36.97 per GB against the Spark's $54.30. The Spark's $6,950 is NVIDIA's Oct. 31 marketplace listing, and its stock was not confirmed.

Against GMKtec's own 128 GB boxes, the step is steep. The EVO-X2 with 128 GB and 1 TB lists at $3,499.99, or $27.34 per GB. The EVO-X5 Pro's price is 94% higher for 50% more memory and about 6.6% more bandwidth. With 2 TB on both, the EVO-X2 costs $3,649.99 and the premium is 86%. GMKtec's EVO-X3, the 128 GB, 2 TB model Notebookcheck compared it with, lists at $3,799.99, a 79% premium. All of these ratios are Voltgrid arithmetic from GMKtec's store prices on Nov. 2.

The launch tiers have ended. GMKtec's Sept. 28 post set three tiers for the 2 TB model: $6,399 for the first 30 units, $6,599 for a seven-day early-bird window, and $6,799 as the regular price. Notebookcheck reported a $6,599 starting price that day. The store now charges the $6,799 regular price, and GMKtec says on an undated, store-wide banner that a price increase is coming soon. It gives no date or amount.

How It Compares on Paper

AMD's own 128 GB box is the clearest guide to what the measurements may show. In Voltgrid's Ryzen AI Halo review, StorageReview measured AMD's $3,999 Halo at 222 tok/s against the DGX Spark's 701 on gpt-oss-120b in vLLM at batch 64. The gap was widest in prefill, where compute matters more than memory. The PRO 495 keeps the same 40 compute units with only a slightly higher clock, so nothing on the spec sheet suggests that gap closes. That is an inference until someone measures it.

Two other machines pair the same chip with 192 GB. Framework says its 192 GB Desktop with the PRO 495 is due to ship in November. Notebookcheck reported that the EVO-X5 Pro starts slightly below Minisforum's MS-S1 Max-P495, which uses the same chip and memory.

For networking, the EVO-X5 Pro's two 10 GbE ports and USB4 v2 links fall well short of the 200 Gbps ConnectX-7 card in every DGX Spark. GMKtec advertises clustering from two units up to a 2U rack but publishes no multi-unit speed figures.

What Is Not Known Yet

The record is missing independent tokens-per-second figures with stated settings. GMKtec's numbers come from its own lab, and it notes that results vary with model, quantization, context length and software. Wall power under a sustained model load is also unmeasured. So is whether 160 GB of GPU memory holds up across Windows and Linux drivers.

The next dates are close. AMD reports third-quarter results on Nov. 3, and Framework's 192 GB Desktop is due to ship in November, which should bring more PRO 495 systems into reviewers' hands. Voltgrid will turn this file into a full review once at least two outlets publish measurements.