Running Local LLMs on 128GB of Memory — GMKtec EVO-X2, Mac Studio and DGX Spark Compared

2026年8月22日

Run an LLM (a large language model) locally and the first wall you hit is memory. The working memory on a consumer graphics card (VRAM, the desk space you spread the model out on) tops out around 24GB, which makes it hard to load anything much beyond 70B (70 billion parameters, the settings inside the model).

So I got curious about the mini PCs and workstations that carry 128GB of memory in a single unit for local AI. This article compares the three machines people talk about — the GMKtec EVO-X2, the Mac Studio (M4 Max) and NVIDIA’s DGX Spark — on memory bandwidth (how fast data moves in and out of memory), power draw, price, software support and generation speed.

None of this is measured on my own hardware. It is a research piece assembled from manufacturers’ published figures, overseas reviews, published benchmarks and Japanese retail prices, with sources given.

The conclusion: as of June 2026 the field of large-memory machines has narrowed sharply, and what you can realistically buy with 128GB comes down to the GMKtec EVO-X2 (cheapest, low power) and the NVIDIA DGX Spark (strong on CUDA). Large-memory Mac Studio configurations (128GB and above) can no longer be ordered because of the DRAM shortage (96GB is the current maximum), and a big-memory Mac means waiting for the M5 Ultra generation. On both value and availability, the EVO-X2 is the front-runner for now.

Prices and performance are published figures, retail prices and review measurements as checked on 15 June 2026. Coupons, exchange rates and stock all move them. Japanese prices are converted at roughly ¥150 to the dollar. Sources are collected at the end.

How the numbers in this article were arrived at

This is a desk study of how far you can get on published information before buying. The figures work as follows.

• Price: Japanese retail after coupons. Memory price inflation affects these, so a check date is attached. Note how much capacity changes the price (128GB against 64GB).
• Memory bandwidth: manufacturers’ published figures (theoretical bandwidth from bus width times memory speed).
• Generation speed: the item needing most care. These are not my measurements; they are quoted from overseas reviews. I then cross-checked them against the principle that local LLM generation speed is roughly proportional to memory bandwidth, to see whether each machine’s figure is consistent with the bandwidth ordering. Reviews sometimes report fast numbers under optimised conditions (rocWMMA and so on), and a plain setup can come in lower. Results from actually running these will be measured separately in a hands-on series.

This article is the pre-purchase comparison; the hands-on series that follows is what happened when the hardware was actually bought and run.

The three machines

All three are compact machines that can carry 128GB.

The GMKtec EVO-X2 is a mini PC built on AMD’s Ryzen AI Max+ 395 (codename Strix Halo). It is an SoC with CPU, GPU and NPU integrated, and its 128GB of LPDDR5X is shared between system and graphics. The 128GB/2TB model is on sale in Japan; as of June 2026, with memory prices up, the street price with coupons is from about ¥520,000 (about $3,470), roughly the same as direct from the manufacturer. The 64GB version is much cheaper at about ¥270,000 ($1,800), and is a strong option if you do not need the capacity.

The Mac Studio (M4 Max) is Apple’s desktop. A warning here: in March 2026 Apple pruned its large-memory configurations, and as of May 2026 the current Mac Studio (2025) offers only 36GB, 64GB and 96GB. 128GB, 256GB and 512GB are gone, reportedly because of the global DRAM shortage. So “a Mac Studio with 128GB" cannot currently be ordered, and 96GB is the ceiling for a Mac (a 70B model at Q4 quantisation — compressing the model to make it smaller, Q4 meaning down to 4 bits — is about 40GB and fits in 96GB, but the very large models discussed below do not). The Mac mini tops out at 64GB with the M4 Pro.

NVIDIA’s DGX Spark is a small AI development machine from the GPU maker. It carries a GB10 Grace Blackwell and 128GB of LPDDR5X.

What you can actually buy right now (June 2026)

Before the specifications, here is availability. Prices and stock move fast, so check before buying.

• GMKtec EVO-X2 (128GB): on sale. From about ¥520,000 ($3,470) with Japanese coupons after the memory price rises, similar to direct sales. The 64GB version is from about ¥270,000 ($1,800). Still the most realistically purchasable 128GB machine for now.
• NVIDIA DGX Spark (128GB): on sale. Japanese reseller pricing is about ¥730,000–760,000 ($4,870–5,070). The option if you want CUDA.
• Mac Studio: large configurations (128GB and above) cannot be ordered; 96GB is the maximum. A big-memory Mac means waiting for the M5 Ultra generation (expected late 2026, delayed).

With the 128GB field temporarily narrowed as of June 2026, the EVO-X2’s relative position has improved.

As an Amazon Associate we earn from qualifying purchases.

Choose the platform first, not the specification

Before the spec sheet, one point matters more than any other. The biggest difference between these three machines (and a discrete GPU machine) is not bandwidth or capacity, but the platform: the OS combined with the GPU compute stack. That governs the decision more than anything else.

Terms first. CUDA is NVIDIA’s GPU compute platform, ROCm is AMD’s equivalent, and Metal and MLX are Apple’s. Whether an AI tool runs at all depends on which of these it supports.

Why is this the biggest difference? Because the GPU stack decides what runs at all, how quickly new models are supported, and how much friction you hit.

• CUDA (NVIDIA): the de facto standard for AI. Nearly every tool supports it, new models arrive immediately, there is more information available than for anything else, and it causes the fewest problems.
• Metal and MLX (Apple): polished, but Apple-only, and some tools lag in support.
• ROCm (AMD): improving, but with the most compatibility friction. “It does not run" and “it needs tweaking" still happen.

The OS matters daily too. Windows is general-purpose and good for games, macOS is refined and creative-leaning, Linux suits development and servers.

The rule is platform first, specifications second.

• (1) Decide the platform from your software requirements and existing setup. For the newest tools, training, image generation and the largest body of information, CUDA (NVIDIA, DGX Spark, or a discrete GPU machine). If you are already on Apple and mostly doing inference, a Mac. If you want large LLMs cheaply and can tolerate ROCm friction, the EVO-X2.
• (2) Then compare bandwidth, capacity and price within that platform.

As a warning: choosing on specifications alone leads to “it should be fast, but the tool I need does not run, and new models take ages to support." Platform fit comes before specification.

Capacity and bandwidth

Running large models depends not only on capacity but on bandwidth, the speed of reading and writing memory. Generation speed, in tokens per second, tends to scale with bandwidth.

All three can carry 128GB, but bandwidth differs. The published figures:

On bandwidth alone the Mac Studio M4 Max stands clear at 546GB/s. The EVO-X2 and DGX Spark are close together in the 250–270GB/s range. There was once a Mac Studio M3 Ultra with up to 512GB at 819GB/s above them, but that configuration can no longer be ordered.

Power, price and platform

The main specifications together. Prices are Japanese retail including tax, and overseas published figures, as of 15 June 2026.

The EVO-X2 is the cheapest. Memory inflation puts 128GB/2TB at about ¥520,000 ($3,470) with coupons, close to direct sales; the 64GB version goes for about ¥270,000 ($1,800). The DGX Spark’s overseas list price starts at $3,999, and in Japan OEM versions (ASUS, Dell and others) and resellers sell at about ¥730,000–760,000 including tax. With the same GB10, NVIDIA’s own Founders model and higher configurations or bundles are reported above ¥1,000,000, so the price band splits. As noted, Mac Studio large configurations cannot be ordered and 96GB is the maximum; the former 128GB (¥583,800) and the M3 Ultra 512GB are unavailable.

On power, the DGX Spark is the highest of the three (a 240W supply, with the GB10 chip itself rated at 140W TDP). The EVO-X2 is in the 140W class in performance mode. The Mac Studio is widely reported as relatively frugal under load, and for a machine that runs constantly the electricity difference is not negligible.

Real generation speeds (from reviews)

Actual speed varies enormously with model, quantisation and backend. Here are measurements from overseas reviews and benchmarks, with sources. These are not my own figures.

Looking at generation speed (tokens per second) for a 70B-class model at Q4:

The GMKtec EVO-X2 is reported at about 11 tokens per second on the Vulkan backend with Llama 3.3 70B at Q4, and about 18 after ROCm optimisation (Hardware Corner and a published Strix Halo testing guide, among others).

The Mac Studio (M4 Max), using its 546GB/s, is reported at roughly 20–30 tokens per second on 70B-class generation, about twice the EVO-X2. It has the advantage in local LLM use, where bandwidth translates directly into generation speed.

The DGX Spark is said to be about level with the EVO-X2 on 70B-class decode, but its distinguishing feature is prompt processing (prefill), which CUDA makes exceptionally fast. LMSYS’s review shows a clear split between prefill and decode at the 70B class (LMSYS, Tom’s Hardware), and reviews such as ServeTheHome’s report prefill reaching thousands of tokens per second under the right conditions (one figure being about 2,817). That matters when you are pushing long prompts through in bulk.

On software maturity, the general view is CUDA first, then macOS’s MLX, with AMD’s ROCm having only recently caught up on tool support. Assume the EVO-X2 means switching between Vulkan and ROCm as needed.

How GPU strength affects local LLMs: prefill and decode

“A strong GPU means a fast LLM" is only half right, because inference splits into two phases with different limiting factors.

The first is prefill. The prompt is read in and prepared in one go, and a large volume of matrix arithmetic runs at once. Raw GPU compute decides the speed here, so a stronger GPU is faster.

The second is decode, generating the answer one word at a time. Producing each word requires reading all the model’s weights from memory, so speed is roughly “memory bandwidth divided by model size." Bandwidth dominates over GPU compute.

So the speed at which text appears (decode) is set mostly by bandwidth, and processing a long input (prefill) is set by GPU strength. That the bottleneck changes between the two phases is shown in review measurements (LMSYS, checked 15 June 2026).

Applied to the three machines:

• Generation (decode): decided by bandwidth, not the GPU. The Mac Studio M4 Max’s 546GB/s comfortably exceeds the 256–273GB/s of the other two, so it generates faster.
• Prefill: raw GPU performance and software optimisation, so the DGX Spark with CUDA stands out and the EVO-X2 is relatively weak.
• Software maturity (CUDA far ahead of MLX, MLX ahead of ROCm) widens the gap, because better-optimised software extracts more from the same hardware.

Which leads to:

• Mostly short conversation and chat: perceived speed is decode, so bandwidth matters and the Mac Studio has the advantage.
• Long documents, RAG (feeding it external documents), code completion, large batch jobs: prefill, GPU performance and CUDA matter, so the DGX Spark has the advantage.
• Running constantly at low power: the EVO-X2.

Large-memory PC or a discrete GPU?

By this point the obvious question is “why not just put a cheap GPU like an RTX 3060 in a PC?" Large unified-memory machines and discrete GPU cards are simply good at different things.

Terms: unified memory is large memory shared by CPU and GPU, a discrete GPU is a card you install in a PC, VRAM is the memory on that card, and quantisation compresses the model to make it smaller (Q4 compresses the weights to 4 bits).

What large-memory PCs are good at

• Capacity lets large models run: 128GB holds a 70B model (about 40GB at Q4) whole. Matching that on discrete GPUs takes two or three 24GB RTX 4090s (over $1,670 each, so over $3,300 together) and a lot of power (450W per card). Getting past the “runs or does not run" wall is the main argument.
• One box: a GPU card is a component and needs a host PC around it. These machines are complete on their own.
• Low power, quiet, small: the EVO-X2’s roughly 140W is far below a multi-card setup (hundreds to a thousand watts), and it takes little space.
• Running large models cheaply: for the goal of “run a 70B," it tends to win on total cost and power against multiple cards.

What they are not good at

• Bandwidth and compute lose to a high-end discrete GPU: within what fits in VRAM, a single strong GPU (the RTX 4090 exceeds 1,000GB/s) is faster.
• Weak on prefill and training: both need compute (the CUDA-equipped DGX Spark is the exception).
• Immature software: ROCm and MLX do not cover as much as CUDA, and some tools need work.
• No upgrades: the memory is soldered to the board. What you buy is what you have.
• Poor for general GPU work: games and video editing favour a strong discrete GPU.

The discrete GPU side

• Good: fast for models that fit in VRAM (broadly up to the 13B class). CUDA is mature with wide tool support, and it is good at training and prefill. Upgradeable, usable for games, and cheap to start with second-hand.
• Not good: 12GB of VRAM on an RTX 3060 will not hold a 70B. Adding capacity means multiple cards, which is expensive and power-hungry. And it needs a host PC.

The capacity wall, laid out:

The conclusion

• If the goal is running a 70B locally: the capacity wall makes a large-memory PC essentially the only route (a single GPU simply runs out of room). Within that, generation speed follows bandwidth (Mac Studio, then DGX Spark and EVO-X2 about level).
• Mostly small and mid-sized models, with training, prefill, existing CUDA work, or gaming: a PC with a discrete GPU suits better.

In a sentence: a large-memory PC to get past the capacity wall, a discrete GPU for speed and versatility within it.

Software and ecosystems

Even for the same job of running an LLM locally, the platform changes which software and quantisation formats are available.

The DGX Spark’s strength is that CUDA works as-is, including FP4 low-precision arithmetic at the hardware level. If you have existing CUDA-based work, it fits.

The Mac Studio runs models through MLX and Metal on macOS. Reports emphasise how comfortably it handles mid-sized models using unified memory and high bandwidth. It has no dedicated FP4 hardware.

The EVO-X2 runs Windows or Linux, mostly using AMD’s ROCm or llama.cpp (Vulkan or ROCm) with 4-bit quantisation such as GGUF Q4. Existing CUDA work does not carry over.

What separates the DGX Spark from the EVO-X2

Both carry 128GB and their inference speeds are close, but underneath they are quite different.

Terms: ARM is the power-efficient CPU design also used in phones (a different lineage from the x86 common in PCs), CUDA is NVIDIA’s GPU compute platform, a PFLOP is a thousand trillion operations per second, and FP4 is 4-bit low-precision arithmetic used to speed up AI.

The DGX Spark is, at heart, an ARM-based Linux machine. The CPU is the GB10 Grace Blackwell, with 20 ARM cores (ten Cortex-X925 performance cores and ten Cortex-A725 efficiency cores, ARMv9). The OS is DGX OS, based on Ubuntu, with the CUDA toolkit and NIM preinstalled, the same distribution as the data-centre DGX machines. The EVO-X2 is an x86 AMD chip with Windows or Linux, much closer to an ordinary PC.

One note: the DGX Spark exists as NVIDIA’s own Founders version and as OEM equivalents (ASUS, Dell, HP, Lenovo, MSI, Acer, Gigabyte and others), but inside they are all the same NVIDIA GB10 Grace Blackwell Superchip. Every OEM machine is ARM, with the same ARM Linux. What differs is the chassis, cooling, ports, storage, brand and price (OEM and resellers at about ¥730,000–760,000, with the Founders version reported around ¥1,200,000). Performance is essentially identical.

That is exactly the same picture as the Strix Halo machines (EVO-X2, Framework, Minisforum and others, all the same AMD x86 chip). Every GB10 machine is the same NVIDIA ARM chip. Whichever DGX-family machine you choose, the platform is ARM Linux plus CUDA, and the difference is price and chassis.

The heart of it: because bandwidth is close (273 against about 256GB/s), token generation speed on a large MoE is about the same on both, roughly 11–13 tokens per second. For simply running an existing model, there is no large speed difference.

What differs decisively is software, compute and purpose. The DGX Spark brings ARM Linux, CUDA, enormous compute (up to 1 PFLOP FP4) and compatibility with data-centre workflows, and it is strong on long-document processing and on training and development. The EVO-X2 is Windows and ROCm, cheap, and inference-oriented.

Bluntly: if you only want to run a large model like Qwen3-235B, the EVO-X2 does it at almost the same speed as the DGX Spark for a third to a quarter of the price. The DGX Spark’s premium buys CUDA, large compute and an ARM Linux development environment. It is for people who build, train and develop against CUDA, and expensive for anyone who just wants to run inference.

GMKtec EVO-X2 (Ryzen AI Max+ 395 / 128GB / 2TB)：Amazon

Coming soon: the RTX Spark (a different thing from the DGX Spark)

Confusingly, there are two Sparks. Separately from the DGX Spark, an RTX Spark has been announced for release in autumn 2026.

Terms: Windows on ARM is the ARM version of Windows, Copilot+ PC is a specification for Windows PCs with enhanced AI processing, and a petaFLOP is a thousand trillion operations per second.

The RTX Spark (formerly codenamed N1X, from NVIDIA and MediaTek) was announced by NVIDIA and Microsoft on 31 May 2026, for release this autumn from ASUS, Dell, HP, Lenovo, Microsoft Surface and MSI. It is positioned as a general-purpose PC, laptop or small desktop. Inside is a 20-core Grace-family ARM CPU, a Blackwell-generation RTX GPU, an NPU and up to 128GB of unified memory, running Windows on ARM (Copilot+ PC). Reporting describes 6,144 CUDA cores, a TSMC 3nm process and petaFLOP-class compute, with Jensen Huang stating that every Windows application runs.

Both are siblings — Grace-family ARM, Blackwell, 128GB unified memory — but differ in OS, purpose and price. The DGX Spark is an ARM Linux and CUDA appliance; the RTX Spark is an ordinary Windows PC.

For local LLMs, the RTX Spark brings Windows, ARM, a Blackwell GPU and up to 128GB into a general-purpose PC, potentially combining the EVO-X2’s Windows compatibility with NVIDIA’s GPU and CUDA strengths. If it arrives, it could redraw the comparison in this article.

For now, though, it is due in autumn 2026, mostly as laptops and small desktops, and CUDA on Windows on ARM and its real local LLM performance remain to be confirmed. I will hold off judging it until prices and performance are settled.

Recommendations by use

Pulling the research together:

To keep costs down and start with large-memory local LLMs at low power, the GMKtec EVO-X2 is the front-runner. It carries 128GB, is available from about ¥270,000 ($1,800) with Japanese coupons in its lower configuration, and at 140W is the most frugal of the three. Running Windows or Linux also makes it easy to get going.

For generation speed and stability, the Mac Studio (M4 Max) would ordinarily be strong on bandwidth, with roughly twice the EVO-X2’s generation speed at the 70B class, plus quiet, low-power operation. But as of June 2026 large configurations cannot be ordered and 96GB is the ceiling (a 70B at Q4 fits; anything larger does not). A big-memory Mac means waiting for the M5 Ultra generation. There was a point when a high-priced EVO-X2 channel (around ¥550,000) closed the gap with a Mac Studio, but with large-memory Macs unavailable, the EVO-X2 is what you buy to have 128GB today.

Finally, if you need fast prefill or have CUDA-based development work, and you are looking towards FP4 and very large models, the NVIDIA DGX Spark is the axis. It is the most expensive and the most power-hungry of the three.

Does any of this apply beyond LLMs?

All of the above concerns large LLMs. For image generation and training, different factors matter. Roughly: only large LLMs have a capacity wall, and image generation and training are the reverse, dominated by GPU compute and CUDA.

Terms: image generation means producing pictures with Stable Diffusion and similar, training and fine-tuning means further training a model to adjust it, and embedding means converting text into numerical vectors.

For image generation, NVIDIA and CUDA are currently the most mature, the fastest, and the most widely supported. SDXL runs practically on an RTX 3060 and quickly on an RTX 4090.

AMD claims that the Radeon 8060S integrated GPU in the Ryzen AI Max+ 395 matches a laptop RTX 4070, and puts it at about 3.9 times a MacBook Pro (M4 Pro) on Stable Diffusion. But those are AMD’s own figures, and ROCm maturity and tool compatibility remain the practical obstacle, so discount accordingly (AMD published figures, from AMD’s developer article and Notebookcheck, checked 15 June 2026).

Image generation runs on a Mac with MLX too, but tends to be slower than an NVIDIA GPU at the same price.

There is an exception to the capacity rule. For large batches, high resolutions and long video that exceed a single GPU’s VRAM, a large-memory PC comes into its own. AMD states the Ryzen AI Max+ 395 can allocate up to 96GB to the GPU (112GB in some environments), which means not stopping for lack of VRAM.

Training and fine-tuning are almost entirely dominated by CUDA and NVIDIA’s ecosystem, so the DGX Spark or an NVIDIA discrete GPU.

As a quick reference:

• GMKtec EVO-X2: for running large LLMs cheaply at low power. Relatively weak on image generation and training.
• Mac Studio (M4 Max): for quiet video editing alongside mid-sized LLMs. Large configurations (128GB+) cannot be ordered; 96GB is the current ceiling.
• NVIDIA DGX Spark: for training, image and video, prefill and development, all through CUDA.
• RTX 3060 (discrete GPU): for image generation, small and mid LLMs, light training and gaming, cheaply.

Who these do not suit

None of them is universal.

The GMKtec EVO-X2 does not suit anyone who wants CUDA-based tools and libraries to run as-is. Assume you will be getting your hands dirty with ROCm, Vulkan and quantisation. Its 70B generation speed is a step behind the Mac Studio.

The NVIDIA DGX Spark is expensive and power-hungry, and since its 70B generation speed is not far from the EVO-X2’s, it is overkill as a way to simply try local LLMs. It suits people with a clear need for long-prompt processing or existing CUDA work.

The Mac Studio (M4 Max) has no dedicated FP4 hardware and will not run CUDA-based software. If you want to standardise on Windows or Linux, or need particular NVIDIA-oriented tools, it is a difficult choice.

One more thing: these integrated large-memory machines are not “maximum performance out of the box." Speed changes with quantisation and backend (Vulkan, ROCm, MLX). Starting small with a graphics card you already have and moving up later is a reasonable approach.

Choosing a GPU for LLMs, by stage

We have established that large LLMs are about capacity, and images and training are about compute and CUDA. So which class of GPU should you actually buy?

The premise: the biggest bottleneck for local LLMs is VRAM. Inference is often decided by bandwidth and capacity, and GPU compute matters less than you would think. “Does it fit" comes first; bandwidth then sets the speed.

By stage (prices as of 2026, and they move):

• Entry, small models (7–13B), image generation, gaming: the x060 class. Cheapest entry is an RTX 3060 12GB; best on budget is an RTX 4060 Ti 16GB (16GB gives headroom and it draws less power). 8GB versions are cramped, so 12GB or 16GB and up.
• Best value, mid-sized models (up to 30B): a used RTX 3090 24GB. 24GB of VRAM at 936GB/s, and the card most guides call the best value per dollar. Used prices run about ¥130,000–180,000 ($870–1,200).
• New high end (RTX 4090 24GB, RTX 5090 32GB): expensive if LLMs are all you want. The compute is more than you can use, 24GB will not hold a 70B, and the capacity wall is the same as mid-range cards (the 5090’s 32GB is comfortable up to the 32B class). Worth it if you also train, generate images and game.
• Large models (70B and up): no single GPU has the capacity. You need a 128GB-class large-memory PC (EVO-X2, DGX Spark; large Macs cannot be ordered as of June 2026) or multiple GPUs.

In short: x060 (12 or 16GB) to start, a used RTX 3090 for the best mid-sized value, a new x090 being expensive for LLMs alone, and a large-memory PC once you pass 70B.

Can this replace cloud AI coding tools?

This may be the real question: buy a 128GB machine and can you replace cloud agentic coding? An honest answer.

Terms: agentic coding is where the AI reads files, writes code and runs tests, repeating the loop by itself. SWE-bench is a leading benchmark measuring whether real software bugs can be fixed. Frontier means the highest-performing models at a given moment.

The conclusion first: as of 2026, a 128GB-class machine cannot yet compete on the hardest agentic coding that these cloud tools are best at. Not because the models are weak, but because there is no local machine with the capacity and speed to run the models that do compete, comfortably.

On models alone, open and closed have converged

2026 benchmark reports put open-weight models close to the frontier. On SWE-bench Verified, open models such as MiniMax, GLM-5, Kimi K2.5 and Qwen 3.5 reach around 76–80%, within a few points of the top closed models. The gap in raw model capability is much smaller than it was.

But those models do not fit in 128GB

Here is the catch. Those near-frontier models are all large (MoE models of 200B to 1T parameters) and do not fit in a 128GB machine. In a mixture-of-experts model only some experts run for each token, but all the weights still have to be in memory. Kimi K2.5, reported at the 1T scale, needs about 280GB even at 4-bit quantisation, which puts it in 512GB territory.

What a 128GB machine runs comfortably is the 70B class through mid-sized MoE models (broadly up to about 120B at 4 bits). That is a step below the near-frontier models, and with slower generation and prefill, agentic coding’s repeated loops do not run at a practical speed.

The former exception can no longer be bought

Until recently the top Mac Studio M3 Ultra (up to 512GB) was the only affordable machine that could run large open MoE models quantised. Since March 2026 Apple has pruned its large configurations, and as of June 2026 not even 128GB can be ordered, let alone 512GB. So running a large MoE whole on accessible hardware is now essentially closed off, until the M5 Ultra generation.

Where to draw the line

In a sentence: the difficult 20% still belongs to the cloud frontier, and the remaining volume runs perfectly well locally at 10 to 100 times lower token cost. The most practical arrangement is a hybrid, running the volume locally and sending the hard parts to the cloud.

In summary: work backwards from what you want to do

In one line: large LLMs are decided by memory capacity, so a large-memory PC; image generation and training are decided by compute and CUDA, so NVIDIA; small models and gaming, a discrete GPU. Choosing backwards from what you want to do is the short route.

What “it runs a 235B" actually means

“A 128GB machine runs a 235B model" sounds startling, so here is what it really is, without the exaggeration.

Terms: a mixture-of-experts (MoE) model holds many experts internally and uses only some of them for each token. Active parameters are the portion actually used. Sparse activation means only a small fraction runs at a time.

What is running is Qwen3-235B (formally Qwen3-235B-A22B), a standard open MoE model, not a custom build. The total is 235B parameters, but only about 22B are active generating each token (picture 8 of 128 experts working). Quantised, it fits in 128GB and runs at about 11 tokens per second. “Enormous yet it runs" is thanks to sparse activation, not magic.

Second, this is not specific to GMKtec. Generation speed is set roughly by bandwidth and the amount of active parameters. So machines with the same chip (AMD Ryzen AI Max+ 395 / Strix Halo) and 128GB of unified memory — the GMKtec EVO-X2, Framework Desktop, Minisforum MS-S1 Max, HP and others — perform about the same on AI (roughly 11 tokens per second on Qwen3-235B). What differs is cooling, expandability, networking, price and shape. A machine with higher bandwidth (the Mac Studio’s 546GB/s against Strix Halo’s 256GB/s) is faster in proportion.

Within that group, the MINISFORUM MS-S1 Max is positioned above the EVO-X2 on expandability. The chip and memory that set AI speed are the same, so generation speed should be about level. The difference is expansion: on published specifications the MS-S1 Max has a PCIe 4.0 x16 slot (running at x4), dual 10GbE, USB4 V2 (80Gbps) and M.2 RAID, while the EVO-X2 has no slot, 2.5GbE and USB4 (40Gbps) at a lower price. Overseas street pricing was higher for the MS-S1 Max (checked July 2026). Expansion and networking point to the MS-S1 Max; price and simplicity to the EVO-X2.

Since this article, the GMKtec EVO-X3 has appeared with the same chip and 128GB. As the parts that determine AI speed are identical, generation speed should be about level with the EVO-X2 (inferred from published specifications; I have not tested an EVO-X3).

To spend less there is the GMKtec EVO-X1 Pro with 64GB. That is a lot of memory for a PC, but not enough for a large local LLM. It does, however, publish OCuLink support (PCIe Gen4 x4), leaving room to add an external graphics card later (published specification, checked 1 August 2026; I have not tested one).

So where is GMKtec’s value? Bluntly: the EVO-X2 is not a magic box. It is the cheapest and most obtainable package of “an AMD chip plus 128GB of unified memory."

• Price and availability: 128GB/2TB at about $2,199 overseas from GMKtec directly (96GB/2TB about $1,800), and from about ¥270,000 ($1,800) in Japan with coupons. The Framework Desktop is $1,639 with 64GB and the same chip (storage and OS extra). Cheap for its class, and the most widely distributed early implementation.
• Form: small, no assembly, a complete machine with power supply and cooling included.

The way to choose: within the Strix Halo machines, choose on price, cooling, expandability and shape; if you purely want speed, choose a higher-bandwidth machine such as a Mac. Treat AI capability as coming from chip plus memory, and GMKtec’s differentiation as price and distribution, and the decision stays clear.

What would that same 235B do on a Mac? (theoretical)

A rough estimate for Qwen3-235B on a Mac Studio (M4 Max), scaled by bandwidth. This is theoretical; I could not find enough published measurements, so treat it as a guide.

Token generation is bandwidth-limited, so bandwidth scales it approximately. The M4 Max’s 546GB/s divided by Strix Halo’s 256GB/s is about 2.1x. Applied to the EVO-X2’s roughly 11 tokens per second on Qwen3-235B, that gives about 23–24, with a realistic expectation of about 20–25 depending on MLX efficiency, quantisation and thermals. Published Mac benchmarks do report figures around 20 tokens per second (sources: MacStories and llmcheck, checked 15 June 2026; theoretical and unverified here).

There is an important caveat. Whether Qwen3-235B fits on a single Mac depends on quantisation: reviews report about 132GB at 4-bit and about 103GB at 3-bit. So the 96GB ceiling on currently orderable Mac Studios will not hold it at usable quality. On the arithmetic a 128GB configuration would fit a 3-bit version, but 128GB and above cannot be ordered. “Take the Mac for speed" holds in theory, but availability of a large-memory Mac is itself the bottleneck.

To summarise: speed means a high-bandwidth Mac (with availability the problem, and the M5 Ultra generation the wait), and price and availability mean the EVO-X2 (with other machines on the same chip performing the same).

Where the GMKtec EVO-X2 stands

In one line: the cheapest, most power-efficient, single-box way into running large local LLMs (the 70B class). Its AI performance comes from the AMD chip plus 128GB, and is about the same as other machines with the same chip. GMKtec’s differentiation is price and availability rather than performance.

Its strengths:

• 128GB for about $1,800–3,670. Matching that with discrete GPUs means multiple expensive cards, and Macs above 128GB cannot be bought at all. “Cheap, now, 128GB" is the practical answer.
• A complete machine that is frugal, small and quiet (about 140W), suited to sitting there running a large LLM continuously.

Its weaknesses, to accept going in:

• Generation speed is about half the Mac Studio M4 Max’s (256GB/s against 546GB/s).
• Prefill and compute are weak.
• ROCm is immature, which hurts on image generation, training and CUDA-based tools.
• It cannot compete on frontier agentic coding.

Positioned against the alternatives:

• Running large LLMs cheaply: the EVO-X2’s own territory
• Comfortable generation and creative work: a Mac (with large memory waiting on the M5 Ultra generation)
• Training, image and video, prefill, CUDA: the NVIDIA DGX Spark
• Small and mid models, images, gaming: a discrete GPU (RTX 3060 to start, a used RTX 3090 for value)

With high-memory Macs unavailable, the EVO-X2’s relative value has risen simply because you can have 128GB cheaply and immediately.

GMKtec EVO-X2 (Ryzen AI Max+ 395 / 128GB / 2TB)：Amazon

Finally, on fit. If you can accept some compromise on speed, versatility and software maturity, and you want a large LLM running cheaply and continuously, the EVO-X2 suits you well. If generation speed, image generation, training, CUDA or the hardest coding matter most, look at a Mac (large memory waiting on the M5 Ultra), a DGX Spark, or a discrete GPU.

Sources (checked 15 June 2026)

• NVIDIA DGX Spark reviews (generation speed, prefill): LMSYS / Tom’s Hardware / ServeTheHome
• DGX Spark hardware specifications: NVIDIA DGX Spark official / Arm Newsroom
• RTX Spark (formerly N1X, due autumn 2026): Tom’s Hardware / NVIDIA Newsroom
• GMKtec EVO-X2 generation speed (70B Q4, Vulkan/ROCm): Hardware Corner
• Mac Studio specifications and the discontinuation of large-memory configurations: 9to5Mac / Macworld
• RTX 3060 and RTX 4090 specifications: TechPowerUp / NanoReview
• Image generation and per-use performance (including AMD’s own figures): AMD developer article (Ryzen AI Max+ 395) / Notebookcheck (Strix Halo analysis)
• GPU buying guides for LLMs: Best GPU for LLM / CoreLab (VRAM per dollar) / Compute Market (3090 against 5060 Ti)
• Qwen3-235B-A22B MoE structure: Qwen3-235B-A22B (Hugging Face) / Qwen official blog
• EVO-X2 pricing and equivalence with other Strix Halo machines: Tom’s Hardware (EVO-X2 review) / ServeTheHome / Level1Techs

Hardware compared here

Prices and stock move, so check the current specifications and prices before buying.

• GMKtec EVO-X2 (AMD Ryzen AI Max+ 395 / 128GB LPDDR5X / 2TB)
• Apple Mac Studio (M4 Max; 128GB and above could not be ordered as of June 2026, with 96GB the maximum)
• NVIDIA DGX Spark (GB10 Grace Blackwell / 128GB LPDDR5X)