That attraction is real, but so are the limits. Shared memory is not the same as having 128 GB of dedicated graphics VRAM, processor TOPS ratings are not a promise of a particular model’s token rate, and a claim that a system can support a 235-billion-parameter model says little without the model format and test conditions. At its listed US price of $3,799.99 at the time reviewed, the EVO-X3 needs to be judged as a specialized compact workstation rather than an inexpensive mini PC.
Already on sale, not a new August launch
GMKtec announced the EVO-X3’s global launch for July 6, 2026, with 128 GB plus 2 TB and 128 GB plus 4 TB storage configurations. That timing matters because later coverage describing it as newly launching in late August would be misleading: the system was already on sale.
The official US product listing showed a $3,799.99 price, although the captured listing did not explicitly connect that price to either the 2 TB or 4 TB option. Prospective buyers should therefore confirm the selected configuration and final checkout price rather than assuming the published figure applies equally to both models.
The physical design also warrants perspective. GMKtec specifies a 353 × 186 × 41 mm metal chassis without its stand and an approximate weight of 2.3 kg. This is compact compared with a tower workstation, but substantially larger than the palm-sized machines often called mini PCs. Its intended home is a desk, studio, lab, or office shelf—not necessarily a travel bag.
The hardware proposition: a high-end APU and 128 GB of shared memory
AMD specifies the Ryzen AI Max+ 395 as a 16-core, 32-thread Zen 5 processor with boost frequencies up to 5.1 GHz and 64 MB of L3 cache. Its integrated Radeon 8060S graphics has 40 graphics cores clocked up to 2.9 GHz; GMKtec describes its implementation as RDNA 3.5. These are ambitious specifications for an all-in-one processor, particularly in a system without a built-in discrete GPU.
Memory is the central differentiator. The EVO-X3 has 128 GB of onboard LPDDR5X-8000 memory connected through a 256-bit interface. AMD lists 128 GB as the processor’s maximum supported memory capacity at that speed, so GMKtec is effectively using the platform at its stated ceiling.
For local AI users, that capacity can be more consequential than raw CPU specifications. Large language models, image-generation tools, large datasets, development environments, virtual machines, and high-resolution media projects all compete for memory. A system that can keep more of a workload in local memory can avoid the severe slowdown that occurs when Windows must rely heavily on storage-based paging.
But the word “onboard” carries an important trade-off: GMKtec says the LPDDR5X memory is not upgradeable. Buyers are choosing the system’s lifetime memory capacity at purchase. The 128 GB configuration is generous today, yet it also means the EVO-X3 cannot later receive a RAM upgrade in the way a conventional desktop workstation often can.
Storage is more flexible. GMKtec specifies two M.2 2280 PCIe 4.0 x4 slots and a stated maximum total SSD capacity of 16 TB. The included 2 TB or 4 TB option may be enough for a typical Windows installation, applications, and current projects, but local-model collections can consume storage surprisingly quickly. The second slot is therefore a meaningful practical advantage for AI experimentation, game libraries, video assets, or local backups.
Why 128 GB does not equal 128 GB of VRAM
The most important qualification for AI buyers is that the EVO-X3 has unified system memory, not a discrete graphics card with 128 GB of isolated VRAM. Windows, applications, the CPU, and integrated graphics all share the same fundamental pool. Memory allocation, graphics drivers, runtime support, model size, context length, and other active software all influence what is usable for a given workload.
AMD’s own Windows documentation provides a more concrete reference point: on a Ryzen AI Max+ 395 system with 128 GB of memory, it describes up to 96 GB of Variable Graphics Memory for Vulkan llama.cpp workloads. That is a substantial allocation for an integrated-graphics platform, but it is not an assertion that every EVO-X3 workload automatically receives 96 GB, nor that all 128 GB becomes graphics memory.
The distinction has immediate consequences. A local model may technically fit only after quantization, with a limited context window, and with few other memory-heavy programs open. Someone expecting the machine to run a large model while simultaneously operating multiple virtual machines, Adobe applications, browser tabs, and development tools may find that the shared-memory advantage has practical boundaries.
GMKtec advertises “Max support for Qwen3 235B.” This should be treated as a compatibility-oriented vendor claim, not as a performance result. The available product material does not identify the model file, quantization level, context size, inference runtime or backend, graphics-memory setting, operating system conditions, or tokens per second. Those details determine whether an experience is merely possible, reasonably responsive, or useful for interactive work.
A 235-billion-parameter model can have sharply different memory needs and output speeds depending on these choices. For prospective buyers, a more meaningful pre-purchase question is not “Does it support this model?” but “Which model variant, at which context length, in which Windows application, and at what measured speed?” Until those answers are supplied, it would be unwise to make that headline claim the sole basis of a purchase.
AI TOPS are capability figures, not application benchmarks
AMD advertises up to 126 overall TOPS for the Ryzen AI Max+ 395, including up to 50 NPU TOPS. These figures describe processor capability under AMD’s rating methodology. They are useful for positioning the chip, especially as Windows software increasingly identifies NPUs as an available acceleration resource.
They do not, however, predict a universal result for local AI. Many popular local-model workflows depend heavily on graphics acceleration and memory bandwidth rather than only the NPU. A given application may use the NPU, Radeon graphics, CPU, or a mixture of all three; another may lack a mature acceleration path for the hardware altogether. A TOPS number cannot by itself establish image-generation time, transcription speed, coding-assistant responsiveness, or language-model token throughput.
For Windows users, software compatibility should be checked before hardware is selected. The relevant question is whether the chosen Windows application and backend support the intended acceleration path, not just whether the PC carries a high AI rating.
Power modes, cooling, and the evidence available so far
GMKtec lists three operating modes: Silent at 54 W, Balanced at 85 W, and Performance with a 140 W peak. AMD separately gives the processor a configurable TDP range of 45 W to 120 W. Those figures should not be merged into a claim that the CPU has a 140 W sustained TDP. GMKtec’s number is explicitly a system-level peak-mode description, while AMD’s specification is the processor’s configurable thermal design range.
This is more than terminology. Power policy can affect long-term clocks, fan behavior, surface temperatures, and electricity consumption. The fast mode may make sense for rendering, compiling, gaming, or batch inference; a quieter profile may be more appropriate for all-day office work or a shared living space. A buyer who needs predictable sustained output should look for testing that compares the specific profiles over extended mixed CPU-and-GPU loads.
Some third-party observations already exist. Notebookcheck reported ETA Prime’s testing of the EVO-X3, including approximately 66 °C during AAA gaming, a maximum of 83 °C in a full-CPU stress test, and roughly 147 W peak power consumption while gaming. These results indicate that independent testing has begun and suggest the cooling solution can keep the system operating under demanding scenarios.
They should still be read carefully. A gaming temperature and a full-CPU stress maximum do not answer every question about acoustics, long-duration combined workloads, room temperature sensitivity, or behavior in each vendor power mode. They also do not establish performance stability for a multi-hour local AI job. The reported gaming power peak is nevertheless a useful reminder that this compact workstation can draw materially more power than a low-wattage mini PC.
GMKtec’s own cooling descriptions contain a smaller but notable inconsistency. Its specifications mention three heat pipes and three cooling fans, while another part of the product material refers to dual ultra-large CPU fans. The available material does not resolve the actual fan arrangement, so the system should not be described with certainty as a three-fan design pending clarification or a teardown.
Ports make it more workstation-like, with one major caveat
The EVO-X3 is well equipped for a fixed Windows desk setup. GMKtec specifies HDMI 2.1, USB4 Type-C at 40 Gbps with video output, Wi-Fi 7, Bluetooth 5.4, and 2.5 GbE networking. That combination supports modern high-speed networking, external storage, displays, and wired office or home-lab deployment.
Its more unusual feature is a rear PCIe Gen4 x4 OCuLink port. GMKtec says an external RTX 40-series or RTX 50-series GPU can be attached using a separate dock. This creates a potentially useful upgrade path: the integrated Radeon graphics and shared-memory design can serve local AI experiments and general work, while an external Nvidia GPU could be added for workloads that benefit from CUDA software support or dedicated VRAM.
There are limitations. The dock is not included, and GMKtec explicitly says OCuLink does not support hot-plugging: the EVO-X3 must be shut down before connecting or disconnecting the eGPU. That makes the setup closer to a semi-permanent expansion arrangement than a convenience peripheral. Independent measurements establishing real-world eGPU performance loss, dock compatibility, and stability were not available in the reviewed material, so broad claims about near-desktop GPU performance would be premature.
Who should consider the EVO-X3?
The EVO-X3 makes the strongest case for a Windows user who specifically values a dense machine with 128 GB of shared high-bandwidth memory, two SSD slots, fast connectivity, and the option of OCuLink expansion. It could suit developers testing local models, researchers working with memory-intensive tools, creators who want a compact desktop, and enthusiasts seeking an AMD APU platform beyond typical integrated-graphics systems.
It is less compelling for buyers who mainly need an affordable office computer, a quiet low-power media box, or the best gaming performance per dollar. The $3,799.99 listed price places it in territory where conventional desktops may offer different combinations of upgradeability, discrete graphics, and storage flexibility. The EVO-X3’s RAM cannot be upgraded, and its headline AI potential depends heavily on the software stack and workload chosen.
Ultimately, the EVO-X3’s 128 GB configuration is meaningful because it expands what can plausibly be attempted on one compact Windows machine. It should not be mistaken for a blanket guarantee that every enormous model will run quickly, comfortably, or at full precision. Buyers should match the system to the exact applications they intend to use, verify the relevant Windows acceleration support, and regard the 235B-model message as a claim requiring workload-specific proof rather than a settled benchmark.
