GMKtec EVO-X2 for local AI: the cheapest 128 GB AMD box (2026)

By /
GMKtec EVO-X2 product page on GMKtec UK, October 2026

The GMKtec EVO-X2 is a compact Windows mini PC built on AMD's Ryzen AI Max+ 395, where the processor and a large built-in graphics chip share one pool of memory. Buy the 128 GB version. Up to 96 GB of it can go to the GPU, which runs mixture-of-experts models around 120B well and fits a 70B model at 8-bit. Expect £2,829.99 on GMKtec UK, plus import VAT on delivery. We rate it 3/5. See our guide to which Mac or mini PC to buy for local AI.

Last updated 22 October 2026. Written by Annie, a search and digital marketing specialist. Based on the maker's specifications, UK prices and published benchmarks; I have not tested this machine myself. See the editorial policy.

GMKtec EVO-X2 at a glance for local AI: £1,729.99 with 64 GB before import VAT, up to 128 GB of memory, runs 70B models, the cheapest 128 GB box in the UK, rated 3 out of 5
GMKtec EVO-X2 at a glance. Prices from UK stores, checked 17 October 2026.

Waiting for a deal? Black Friday is 27 November 2026. Verified UK price drops on the AI mini PCs, and what to check before buying from a maker's store, are on the Black Friday AI deals page.

Key facts

  • What it is: A small desktop mini PC from Chinese maker GMKtec, sold in the UK through its own web shop and posted from mainland China.
  • Chip: AMD Ryzen AI Max+ 395 with 16 CPU cores, Radeon 8060S graphics (40 compute units) and a 50 TOPS NPU.
  • Memory options: 64 GB or 128 GB of LPDDR5X-8000. It is soldered to the board, so you cannot add more later.
  • Usable for AI: up to 96 GB of the 128 GB (48 GB of the 64 GB) can be set aside as graphics memory in AMD's Adrenalin software or the BIOS. On Linux you raise the shared GPU limit instead.
  • Memory bandwidth: 256 GB/s
  • UK price: from £1,729.99 (64 GB, 1 TB), as checked 17 October 2026. The 128 GB model is £2,829.99. GMKtec's UK shop lists these prices without VAT and posts from China, so expect to pay import VAT on delivery. Check the total at checkout.
  • Where to buy: GMKtec UK, shipped from China in 7-14 working days.
  • Best for: buyers who want 128 GB of AI memory for the least cash and are happy to deal with an import and a one-year warranty.
  • Our rating (desk review, not hands-on): ★★★☆☆ 3/5

On this page

Which AI models can the GMKtec EVO-X2 run?

With 128 GB it holds any model up to about 150 billion parameters at 4-bit, by our maths of about 0.6 GB per billion parameters.

Memory optionBiggest model at 4-bitExample modelsPublished speed
64 GB (about 48 GB for the GPU)About 70B, with short conversations onlyQwen3 32B (19.4 GB), Llama 3.3 70B (42.4 GB), DeepSeek-R1 32B (GMKtec's own example)No published figure for the 64 GB model
128 GB (up to 96 GB for the GPU)About 150BLlama 3.3 70B at 4-bit or 8-bit (about 74 GB), gpt-oss-120b, Qwen3.5-122B-A10BLlama 3.3 70B Q6_K: 3.7-3.8 tokens a second; Qwen3 32B: 10.4-10.5 tokens a second (ServeTheHome)

Allow 2-4 GB more for a long conversation, which is why a 70B model is a squeeze on the 64 GB version. Mixture-of-experts models read only part of their weights for each word, so they run several times faster on this chip than dense models of a similar size. The big Qwen3 235B-A22B needs 142.5 GB at 4-bit and does not fit.

How fast is it?

ServeTheHome is the main published test of the EVO-X2 itself, on a 128 GB unit:

  • Llama 3.3 70B Instruct, Q6_K, fully on the GPU: 3.7-3.8 tokens a second. Split between CPU and GPU it fell to about 2 tokens a second.
  • Qwen3 32B: 10.4-10.5 tokens a second. ServeTheHome did not state the quantisation or the app.

The same chip has newer, more detailed figures from Framework, measured on its Framework Desktop in August 2026 with llama.cpp, a 2,048-token prompt and 128 tokens generated:

  • Qwen3.8 27B, UD-Q4_K_XL: 11.59 tokens a second with ROCm 7.14 and 12.09 with Vulkan.
  • Qwen3.6 35B-A3B (mixture of experts), UD-Q4_K_XL, ROCm: 52.13 tokens a second.
  • DeepSeek V4 Flash 0731, UD-IQ2_XXS, ROCm: 14.90 tokens a second.

For comparison, ServeTheHome measured gpt-oss 120B in LM Studio at 31.41 tokens a second on the Beelink GTR9 Pro and about 45 tokens a second on the Framework Desktop, both with this chip. Nobody has published a gpt-oss figure for the EVO-X2. Framework notes that the 256 GB/s memory bandwidth is what limits text speed on all of these machines.

Which version should you buy?

Buy the 128 GB model with a 1 TB drive at £2,829.99. It is the reason to own this machine: 96 GB for the GPU holds 70B models at high quality and the fast 100B-plus mixture-of-experts models that 64 GB cannot.

Pick the 64 GB model at £1,729.99 only if you will stay with models of 32B and below. The 2 TB version costs £190 more, but there is a second M.2 slot, so you can add your own drive later. GMKtec now also sells the newer EVO-X3 with the same chip at £3,029.99 for 128 GB, so the EVO-X2 remains the cheaper of the two.

UK prices and where to buy

VersionMemoryStorageUK priceWhere to buy
EVO-X264 GB1 TB£1,729.99, VAT not includedGMKtec UK
EVO-X2128 GB1 TB£2,829.99, VAT not includedGMKtec UK
EVO-X2128 GB2 TB£3,019.99, VAT not includedGMKtec UK

GMKtec shows these as sale prices against £2,029.99, £3,629.99 and £3,849.99, with a free keyboard and mouse while stock lasts. I found no refurbished or education pricing, and no UK high-street retailer stocking it.

Hidden costs

  • Import VAT and courier fees: GMKtec's UK shop does not add VAT and posts from China. GOV.UK says VAT on goods over £135 is paid to the courier before delivery. At 20% that is about £346 on the 64 GB model and £566 on the 128 GB, plus any courier handling fee. Customs duty on computers is 0%.
  • Memory you cannot add later: the LPDDR5X is soldered, so buy the size you will need in three years.
  • Storage: 1 TB fills quickly when a 70B model is 42.4 GB at 4-bit. There is a second M.2 slot for up to 8 TB.
  • Warranty and returns: only a one-year limited warranty, and a 15% restocking fee on change-of-mind returns.
  • Power: ServeTheHome measured 147-160 W while running a 70B model.

Software: Ollama, LM Studio and the rest

The EVO-X2 ships with Windows 11 Pro, and GMKtec says Ubuntu works too. There is no NVIDIA CUDA and no Apple MLX, so everything runs through llama.cpp on AMD's ROCm or the Vulkan graphics interface.

  • llama.cpp: both ROCm and Vulkan backends work on this chip, and in Framework's tests neither is always faster.
  • Ollama: on Linux it supports the Ryzen AI Max+ 395 through ROCm 7. On Windows the chip is not on Ollama's ROCm list, so it uses Vulkan, which is on by default. See our Ollama guide.
  • LM Studio: offers both a Vulkan and a ROCm llama.cpp runtime on Windows. Set GPU offload to maximum when you load a model. See our LM Studio guide.
  • Setup step on Windows: open AMD Software: Adrenalin Edition, go to Performance, then Tuning, and set Variable Graphics Memory, then restart. AMD suggests its Custom (75%) setting for AI work.
  • Setup step on Linux: AMD recommends keeping the BIOS reservation small and raising the shared limit with its amd-ttm tool. ROCm needs a recent kernel: 6.18.4 or later on most distributions, or Fedora 43, Ubuntu 26.04 or Ubuntu 24.04 with the latest HWE kernel.

Noise, power and size

  • 193 x 185.8 x 77 mm, with an external 230 W power adapter (19.5 V, 11.8 A).
  • GMKtec rates the chip at 120 W sustained and 140 W at peak. A front button switches between quiet, balanced and performance modes.
  • ServeTheHome measured 8-14 W and 37-38 dBA at idle in a studio with a 34 dBA noise floor.
  • Running Llama 3.3 70B it drew 147-160 W at 41-43 dBA. ServeTheHome called it far from loud but wished it were quieter.

Who makes the GMKtec EVO-X2?

  • Shenzhen GMK Technology Co., Ltd., founded in 2019, with a factory in Longhua, Shenzhen and an office in Kwun Tong, Hong Kong.
  • Its first mini PC, the NucBox KB1, launched through crowdfunding in 2020. It now sells in more than 70 countries.
  • UK orders, returns and warranty claims go to GMKtec by email (service@gmktec.com). It replies in one to two working days, on China time.
  • There is no UK warehouse. Warranty repairs go to whichever service location GMKtec names, and you may pay the postage.

What's good

  • Cheaper for 128 GB than the Minisforum, Beelink or Framework, even after import VAT.
  • Up to 96 GB for the GPU: room for 70B at 8-bit or 100B-plus mixture-of-experts models.
  • Very low idle power, 8-14 W in ServeTheHome's test.
  • Two M.2 slots for up to 16 TB of storage.
  • Two USB4 ports, HDMI 2.1 and DisplayPort 1.4 for up to four screens.

What's not so good

  • Ships from China without VAT in the price, so the price you see is not the price you pay.
  • One-year warranty, a seven-day cancellation window and a 15% restocking fee on change-of-mind returns.
  • Only 2.5 Gb Ethernet, where the Beelink and Minisforum have dual 10 Gb.
  • Louder under AI load than the Beelink GTR9 Pro in ServeTheHome's tests (41-43 dBA against 39-41 dBA).
  • The spec sheet is muddled, listing 8000 and 8533 MT/s memory and configurations not sold in the UK.

Who should buy it?

Buy the EVO-X2 if you want the most AI memory for the least money and are comfortable paying import VAT at the door. It suits a home lab or a tinkerer who runs Windows or Linux and does not need fast networking. If you want a UK seller, VAT included and a longer warranty, the Beelink GTR9 Pro or the Framework Desktop is the safer buy. If you need NVIDIA's CUDA software, look at the NVIDIA DGX Spark instead.

Buying in the UK

  • Where it comes from: GMKtec UK dispatches from mainland China, usually in one to two working days, then 7-14 working days by DHL, UPS or a local courier. There is no UK warehouse.
  • VAT and duty: GMKtec does not charge VAT at checkout and its policy makes customs charges your responsibility unless it agrees otherwise in writing. Under GOV.UK rules you pay 20% VAT to the courier on goods over £135. The UK tariff for computers (commodity 8471 50 00) carries 0% duty. Keep the receipt and contact GMKtec within seven days of paying if you think it should reimburse you.
  • Warranty: a one-year limited warranty from the date of order, handled by GMKtec. Your statutory rights still apply.
  • Returns: you can cancel within seven days of delivery. Change-of-mind returns lose 15% of the refund and you pay the postage back.

Interesting facts about the GMKtec EVO-X2

  • GMKtec says its EVO series launched in 2025 with an endorsement and signatures from AMD's chief executive, Lisa Su.
  • The EVO-X2 was the first Ryzen AI Max+ 395 system ServeTheHome reviewed. It is designed to stand upright, though the port labels read sideways.
  • Running Llama 3.3 70B on the GPU needed over 50 GB of graphics memory, more than NVIDIA's L40S data-centre card has, as ServeTheHome pointed out.

Alternatives to the GMKtec EVO-X2

  • Beelink GTR9 Pro - pick it for a UK price with VAT included, a three-year warranty and dual 10 Gb Ethernet.
  • Minisforum MS-S1 MAX - pick it for UK warehouse stock, a PCIe slot and a 192 GB option.
  • Framework Desktop - pick it for a UK company, 30-day returns and some of the quietest cooling of the AMD boxes.

Our rating

AreaScore
Memory for AI4/5
Speed3/5
Value4/5
Software3/5
UK buying2/5
Overall★★★☆☆ 3/5

The EVO-X2 gives you 128 GB of AI memory for less than any rival. Import VAT, a one-year warranty and returns to China pull it down.

Frequently asked questions

Can the GMKtec EVO-X2 run Llama 70B?

Yes. ServeTheHome ran Llama 3.3 70B at Q6_K entirely on the GPU of a 128 GB EVO-X2 at 3.7-3.8 tokens a second. The 64 GB model can just hold a 4-bit 70B (42.4 GB) in its 48 GB of graphics memory, with little room for a long chat.

How much RAM does the GMKtec EVO-X2 need for AI?

Get 128 GB. It lets you give 96 GB to the GPU, which covers 70B models and the fast 100B-plus mixture-of-experts models. The 64 GB version is fine for 32B models and below.

Is the GMKtec EVO-X2 good for AI?

It is good value for running large models privately at home, at reading speed or a little slower. It is not fast for dense 70B models, and it cannot run CUDA-only software.

Sources

More guides

Scroll to Top