GMKtec's Evo-X5 Pro is the first "Gorgon Halo" mini PC on sale: AMD's Ryzen AI Max+ PRO 495 with up to 192 GB of unified memory, from $6,799. The headline spec is the memory: 192 GB of LPDDR5X-8533 shared between CPU and GPU, with up to 160 GB allocatable as graphics memory. GMKtec's headline claim is offline LLMs up to 320B parameters.

Every other site will repeat that press release. Here is what they skip: the per-GB price is roughly double a Strix Halo box, only two of the three M.2 slots run at full x4 speed, and there is no barebones option. This guide covers what fits, what it costs, and when to pick it.

Correction note (13 Oct 2026): GMKtec's store page loaded without its full spec sheet at writing time, so some details below (fan layout, exact port list) come from press coverage and reviewer teardowns, cited per claim.

Source note: specs, prices and scores below come from GMKtec's official launch blog and store page (Septโ€“Oct 2026) plus Tom's Hardware, TechRadar and VideoCardz coverage. They are vendor-reported, not an independent VelsTech Solutions benchmark. Street prices move with early-bird tiers โ€“ check the live store page before buying.

Related: our Linux mini PC buying guide (where the Evo-X5 Pro sits as the flagship AI tier) and the local-AI desktop showdown.

Illustration of a compact AI mini PC with memory modules
Illustration of a 192GB-class local-AI mini PC โ€“ not a photo of the actual Evo-X5 Pro.

The short version: Evo-X5 Pro in 60 seconds

What the Evo-X5 Pro actually is

Same idea as last year's Strix Halo mini PCs, changed memory ceiling: a compact desktop whose CPU, GPU and NPU all drink from one big 192 GB pool instead of a small VRAM cup plus system RAM. What stayed identical is the formula (AMD APU + unified LPDDR5X + mini-PC chassis); what is genuinely new is the PRO 495 chip with 16 full Zen 5 cores and the jump from 128 GB to 192 GB.

Why now: 70B-class models at usable quants need ~50โ€“90 GB of shared memory, which 128 GB Strix Halo boxes already cover โ€“ but 100B+ and 300B-class experiments don't fit, and cloud H100 time rents by the hour. AMD's Gorgon Halo answers with a single socket that holds the whole model, and GMKtec, Framework and Minisforum are racing to ship it first.

Specs at a glance

FeatureEvo-X5 ProStrix Halo box (e.g. EVO-X2)
CPU16ร— Zen 5, up to 5.2 GHz12ร— Zen 5 (AI Max+ 395)
GPURadeon 8065S, 40 CUs, RDNA 3.5Radeon 8060S, 40 CUs, RDNA 3.5
NPUXDNA 2, up to 55 TOPSXDNA 2, up to 50 TOPS
Unified memory192 GB LPDDR5X-8533up to 128 GB LPDDR5X-8000
Max shared VRAMup to 160 GBup to 96 GB

The honest row is the power slider: silent and balanced modes hold the chip to 54W and 85W, while performance mode allows 120W sustained and 160W peaks. Expect the full 160 GB VRAM figure and the biggest-model claims to assume performance mode with the vapor-chamber cooler working hard โ€“ in silent mode, clocks (and tokens/sec) will be lower.

Pricing: what it actually costs

ConfigEvo-X5 ProEVO-X2 (Strix Halo)Framework Desktop 495
192 GB + 2 TB$6,799โ€“from $6,799
192 GB + 4 TB$7,099$2,199 (128 GB + 2 TB)192 GB configs similar money

How it compares

What mattersEvo-X5 Pro (vendor/reviewer)Strix Halo 128 GB box
Fits 70B Q4 comfortablyyes, with headroomyes
Fits 100B+ / 300B-classclaimed (320B offline)no โ€“ 128 GB ceiling
Price per GB~$35/GB~$17โ€“22/GB

Read honestly: the Evo-X5 Pro wins only where capacity is the bottleneck โ€“ models that simply do not fit in 128 GB. Everywhere else it trails on value: a Strix Halo box runs the same 70B models for roughly half the money per GB. Treat the 320B claim as marketing until independent reviewers publish tokens/sec on a named quant.

The India angle

Who should buy it โ€“ and who should skip it

What this means if you run local AI

From the real-hardware desk: what changed is the ceiling โ€“ 192 GB unified plus 160 GB of addressable VRAM moves 100B-class experiments from "rent an H100" to "a box under a desk". What didn't change is the workflow: you still pick a quant that fits, you still budget VRAM first (see the VRAM planner and GGUF size calculator), and AMD's ROCm/Vulkan stack on Linux still trails CUDA for exotic tooling โ€“ check the ROCm vs Vulkan guide before committing.

Bottom line

Not a budget mini PC but the current capacity king: if your models fit in 128 GB, buy Strix Halo and pocket the difference. If they don't โ€“ 100B-plus local work, distilling, or serving big quants offline โ€“ the Evo-X5 Pro is the cheapest single-box way to hold them in one machine. Wait if you can: second-wave Gorgon Halo boxes (Minisforum, more GMKtec tiers) should sharpen this pricing within quarters, not years.

Sources

FAQ

What is the GMKtec Evo-X5 Pro?

A mini PC built around AMD's Ryzen AI Max+ PRO 495 with up to 192 GB of LPDDR5X unified memory, up to 160 GB of it allocatable as graphics memory. GMKtec positions it as an "agentic AI PC" for large local models, claiming offline LLMs up to 320B parameters.

How much does the Evo-X5 Pro cost, and is it available?

$6,799 for 192 GB RAM + 2 TB SSD and $7,099 for 192 GB + 4 TB. It went on sale in late September 2026, the first 100 units sold out, and GMKtec's US store showed stock at writing time (mid-October 2026).

Can I run local AI on it, and can I buy it in India?

Yes โ€“ that is its whole point: up to 160 GB of shared VRAM for large local models, with AMD's open Linux drivers. There is no India storefront, so budget roughly โ‚น7 lakh+ landed after shipping, GST and duty, and import it via GMKtec US.