Core Ultra X7 358H with 64GB soldered LPDDR5X-8533, Arc B390, dual 10GbE + 2.5GbE and OCuLink. The site’s first Intel box. Half the memory and half the bus of a 128GB Halo cube.
Street price (USD)
$2,399.99–$2,599.99
Memory
64GB LPDDR5X-8533 (onboard)
Bandwidth
~136 GB/s theoretical (not published by GMKtec)
Chip
Intel Core Ultra X7 358H
Power
35 / 45 / 54W modes, 60W peak; 19V 7.8A DC in
Catch
64GB ceiling; soldered; not CUDA
What it is
GMKtec’s Panther Lake mini. Intel Core Ultra X7 358H on Intel 18A: 16 cores / 16 threads (4 P-cores to 4.8 GHz, 8 E-cores to 3.5 GHz, 4 low-power E-cores to 3.3 GHz), 18MB cache. Arc B390 iGPU with 12 Xe-cores at up to 2.5 GHz, rated 122 TOPS INT8, plus an NPU rated up to 50 TOPS. 64GB onboard LPDDR5X-8533. Two M.2 2280 slots (one PCIe 5.0 ×4, one PCIe 4.0 ×1). Windows 11 Pro preinstalled; GMKtec lists Ubuntu / Linux support. Chassis is 154 × 151 × 73.6 mm, about 2.3 kg, vapor chamber plus dual fans.
Who it is for
People who want a quiet, shipping x86 desk box for 8B–32B-class local models, OpenVINO / Intel stacks, and fast networking, at about a third of the ticket of this line’s 192GB EVO-X5 Pro. Wrong if you need 70B at a usable speed, CUDA, or more than 64GB ever: the memory is soldered and there is one SKU.
What it actually runs
Comfortable: 8B–14B, and quantized 24B–32B dense, on the Arc iGPU or CPU with room left for context.
Possible: a quantized 70B fits in 64GB, but at ~136 GB/s theoretical it is a slow read, not a daily driver. MoE models with small active parameters fare better than dense 70B.
Not the win: anything over ~50GB of weights, training, CUDA-only stacks, or matching the 128GB EVO-X2 / Spark class on capacity.
Standout details
Price snapshot 26 Oct 2026, GMKtec US store: 64GB + 1TB PCIe 4.0 SSD $2,399.99; 64GB + 853GB Phison aiDAPTIV+ AI SSD (PCIe 5.0) $2,599.99. Both listed in stock. The page carries a “price increase coming soon” banner.
Only the Ultra X7 358H is sold on the page today, although the product title also names the X9 388H.
Bandwidth: Intel ARK lists two memory channels; at the 8533 MT/s GMKtec ships, a 128-bit LPDDR5X bus works out to ~136 GB/s theoretical. GMKtec publishes no figure. That is roughly half the ~256 GB/s of the Strix Halo cubes, and decode speed follows bandwidth.
The AI SSD is Phison aiDAPTIV+ (768GB + 85GB). GMKtec’s “up to 40x faster local LLM performance” is vendor copy, not something we measured; it is not extra memory bandwidth.
Rear: OCuLink (PCIe 4.0 ×4, no hot swap) for an eGPU, 10G + 2.5G RJ45, HDMI 2.1, DP 1.4, USB4. Front: a second USB4 (both USB4 ports accept 100W PD input), 2× USB 3.2 Gen2, USB 2.0, audio. Wi-Fi 7 and Bluetooth 5.4.
Intel rates the 358H at 25W base, 80W max turbo; GMKtec runs it at 35 / 45 / 54W modes with a 60W peak.
One-year warranty and a 7-day return window on the GMKtec store.
Honest verdict
Buy the EVO-T2 if you want a shipping Intel mini with 64GB, dual 10GbE + 2.5GbE and an OCuLink escape hatch, and 8B–32B models are the job. Skip the $200 AI SSD unless you have a workload for aiDAPTIV+. If 70B is the job, the 128GB EVO-X2 or Framework Desktop has twice the memory and roughly twice the bandwidth. If you need CUDA, shop Spark.
Prices and stock move. Affiliate links stay off until this directory has real traffic.