GMKtec EVO-X2: review for local AI, 64 or 128 GB

The GMKtec EVO-X2 is a Ryzen AI Max+ 395 mini PC whose soldered LPDDR5X memory (64 or 128 GB) is also used by the Radeon 8060S GPU. For local AI, the 128 GB version is the right choice: up to 96 GB for the GPU, enough to load gpt-oss 120B or Llama 4 Scout. On November 18, 2026, gmktec.fr lists it at €3,499.99 (128 GB + 2 TB) versus €2,029.99 for the 64 GB version.
The GMKtec EVO-X2 is one of the most frequently mentioned mini PCs for running large language models at home. This guide brings together the official spec sheet, public prices as of November 18, 2026, reviews from ServeTheHome and Tom's Hardware, and the settings documented by AMD. You'll find out whether it fits your use case, which memory configuration to choose, and how to set it up on Windows or Linux.
Choosing a machine? Our picks by budget → · Our spec sheet Ryzen AI Max mini-PC with 128 GB →
Good value for local AI: a GMKtec EVO-X2 — Amazon may show another configuration; check before buying.
A mini PC is a complete machine: check the available memory and engine compatibility. It does not replace macOS/MLX or CUDA.
Compare all options by budget, from €800 to €3,500 →
Small budget: RTX 5060 · Large models: RTX 5090 · Mac Studio.
On the go: which laptop for local AI →
Affiliate links — commission possible at no extra cost to you. As an Amazon Associate, BestLLMfor earns from qualifying purchases.
Amazon may redirect you to a local store or show another configuration. Check the listing before buying; this link does not guarantee the destination or configuration.
#GMKtec EVO-X2: is it good for local AI?
The GMKtec EVO-X2 is a good choice for local AI if you want to load models with 60 to 90 GB of weights in a compact case, on Windows or Linux, and moderate speed is enough for you. It isn't a fast machine: it's a machine that offers a lot of graphics memory in a small case.
Unified memory is a single pool of memory shared by the CPU and the integrated GPU: with no separate VRAM, a 70 GB model fits in graphics memory, whereas a consumer graphics card tops out at 16 or 32 GB. According to AMD, up to 96 GB of the Ryzen AI Max+ 395's 128 GB can be allocated as VRAM on Windows.
The trade-off is bandwidth: AMD specifies 256 GB/s, far less than a high-end graphics card. A large dense model therefore runs slowly on the EVO-X2, while a mixture-of-experts (MoE) model, which activates only part of its weights for each token, remains comfortable to use.
- Who it's for
- Developers, freelancers, and small businesses that want a large local model (gpt-oss 120B, Llama 4 Scout) without a discrete graphics card.
- Who it's not for
- Anyone who wants maximum speed on small models, depends on CUDA, or worries about fan noise.
#What's inside the GMKtec EVO-X2? The specs that matter for AI
The GMKtec EVO-X2 combines a 16-core, 32-thread AMD Ryzen AI Max+ 395 processor, a Radeon 8060S integrated GPU with 40 RDNA 3.5 compute units, and 64 or 128 GB of soldered LPDDR5X. The gmktec.fr product page, accessed November 18, 2026, lists stable operation at 120 W, with peaks of 140 W.
| Item | Published value | What this means for an LLM |
|---|---|---|
| Processeur | Ryzen AI Max+ 395, 16 cores / 32 threads | Sufficient for data preparation and CPU computation |
| Integrated GPU | Radeon 8060S, 40 RDNA 3.5 cores | Runs inference via ROCm or Vulkan |
| Memory | 64 or 128 GB LPDDR5X 8000 MHz, soldered | Sets the maximum model size, with no upgrade path |
| Memory bus | 256-bit (eight 32-bit controllers, according to GMKtec) | 256 GB/s bandwidth according to AMD: limits speed |
| Graphics memory | Up to 96 GB on the 128 GB version (AMD, GMKtec) | Lets you load a 70 to 90 GB model |
| Power | 120 W sustained, peaks of 140 W; external power brick of about 230 W | Modest power draw for this capacity |
| Format | 193 x 185.8 x 77 mm; two M.2 slots, up to 16 TB | Fits on a desk, stores several large models |
#How much does the GMKtec EVO-X2 cost in November 2026?
As of November 18, 2026, the official gmktec.fr store lists the GMKtec EVO-X2 at €2,029.99 for 64 GB + 1 TB and €3,499.99 for 128 GB + 2 TB. The struck-through prices are €2,299.99 and €3,599.99, respectively. The gap between the two versions is therefore €1,470.
| Model and configuration | Displayed price | Crossed-out price |
|---|---|---|
| EVO-X2, 64 GB + 1 TB | 2 029,99 € | 2 299,99 € |
| EVO-X2, 128 GB + 2 TB | 3 499,99 € | 3 599,99 € |
| EVO-X3, 128 GB + 2 TB (same processor, OCuLink port) | 3 599,99 € | not listed |
| EVO-X5 Pro, 192 GB + 2 TB (Ryzen AI Max+ PRO 495) | 6 599,00 € | not listed |
GMKtec states on the same page that its prices are expected to rise in 2027 due to component costs, and offers a price guarantee until December 31, 2026. These prices change often during year-end promotions: check them on the official product page on the day you buy.
#GMKtec EVO-X2 64 GB or 128 GB: which version should you choose?
For local AI, choose the 128 GB version of the GMKtec EVO-X2, unless you're only targeting models under 30 GB. The memory is soldered and GMKtec states it isn't upgradable: the choice you make at purchase is final, and what sets the EVO-X2 apart from a regular PC is precisely its large capacity.
GMKtec itself states the difference: DeepSeek-R1 32B for the 64 GB version, and DeepSeek-R1 70B and Llama 4 109B for the 128 GB version. On Linux, AMD's ROCm documentation states that the memory limit shared with the GPU defaults to about 50% of total memory, or about 32 GB on the 64 GB version until you raise it.
| Model | Weights in Q4 | EVO-X2 64 GB | EVO-X2 128 GB |
|---|---|---|---|
| gpt-oss 20B | 13 GB | Yes | Yes |
| Gemma 4 31B | 18 GB | Yes | Yes |
| Qwen 3.6 35B-A3B (MoE) | 21 GB | Yes | Yes |
| Qwen3-Coder-Next 80B-A3B (MoE) | 48 GB | Borderline, short context (estimate) | Yes |
| Llama 4 Scout 109B (MoE) | 65 GB | No | Yes |
| gpt-oss 120B (MoE) | 70 GB | No | Yes |
| Mistral Small 4 119B (MoE) | 72 GB | No | Yes, moderate context |
| GLM 5.3 Flash 320B-A18B (MoE) | 186 GB | No | No |
The 64 GB column is based on an estimate: after reserving 12 to 16 GB for the system and applications, about 48 GB remain for the model and its context. But the context comes on top of the weights, and a long document can take several more gigabytes.
#How fast do LLMs run on the GMKtec EVO-X2?
For the GMKtec EVO-X2, GMKtec claims 13.83 to 57.15 tokens per second for MoE models, while ServeTheHome measured under 4 tokens per second for a dense 70-billion-parameter model. The available figures come from the manufacturer, AMD, and ServeTheHome; none were measured for this guide.
| Model | Throughput | Source and status |
|---|---|---|
| gpt-oss 20B | 57.15 tokens/s | Announced by GMKtec, 128 GB version |
| Qwen3 30B | 55.02 tokens/s | Announced by GMKtec, 128 GB version |
| gpt-oss 120B | 19.25 tokens/s | Announced by GMKtec, 128 GB version |
| Llama 4 Scout 109B Q4_K_M | 13.83 tokens/s | Claimed by GMKtec; AMD claims up to 15 tokens/s |
| DeepSeek-R1 32B (dense) | 9.26 tokens/s | Announced by GMKtec, 128 GB version |
| Llama 3.3 70B Q6_K (dense), fully on GPU | 3.7 to 3.8 tokens/s | Measured by ServeTheHome, review published July 11, 2025 |
The theoretical generation ceiling is roughly the bandwidth divided by the size of the weights read for each token. For a dense 70B model in Q4, about 40 GB, the estimate comes to 256 ÷ 40 ≈ 6.4 tokens per second at best. In Q6_K, this model takes up more than 50 GB according to ServeTheHome: the estimated ceiling drops to around 5 tokens per second, and the measured 3.7 to 3.8 stays below that.
With an MoE model, only the active experts are read for each token, which explains the 19.25 tokens per second reported for gpt-oss 120B despite 70 GB of weights. On the EVO-X2, favor gpt-oss 120B, Qwen 3.6 35B-A3B, or Qwen3-Coder-Next over a dense 70B model.
#How do you configure the GMKtec EVO-X2 for local AI on Windows or Linux?
The GMKtec EVO-X2 ships with Windows 11 Pro, and GMKtec states it is compatible with Linux and Ubuntu. On Windows, increase the graphics memory in the AMD driver; on Linux, install a recent kernel and raise the limit on memory shared with the GPU.
#On Windows: Variable Graphics Memory and LM Studio
- 01Update the AMD driverInstall the latest version of AMD Software: Adrenalin Edition. On July 29, 2025, AMD announced that the Adrenalin 25.8.1 driver lets you load models with up to 128 billion parameters in llama.cpp Vulkan on Windows.
- 02Increase graphics memoryIn AMD Software: Adrenalin Edition, open the Performance tab, then Tuning, and set Variable Graphics Memory, up to 96 GB on the 128 GB version. Restart if the software prompts you to.
- 03Configure LM StudioIn LM Studio, check manual parameter selection and set GPU Offload to the maximum, as recommended by AMD, which also advises Q4_K_M quantization for everyday use.
- 04Verify that the GPU is workingDuring generation, open Task Manager and go to the Performance tab: GPU memory should hold the model and the GPU should be active. If only the CPU is working, the model isn't loaded on the GPU.
For Ollama, the official documentation lists the Ryzen AI Max+ 395 among the chips supported by ROCm v7 on Linux, but not in its ROCm list for Windows, which is limited to Radeon RX 7000 and Radeon PRO. On the EVO-X2 under Windows, Ollama therefore relies on its Vulkan engine, which the same documentation says is enabled by default and already included with most Windows drivers.
#On Linux: recent kernel and shared memory limit
AMD's ROCm documentation requires Linux kernel 6.18.4 or later for Ryzen AI Max, or on Ubuntu 24.04 the HWE kernel 6.17.0-19.19~24.04.2 or later. AMD recommends keeping a small VRAM reservation in the BIOS, such as 0.5 GB, and instead raising the shared memory limit with the amd-ttm tool.
Running amd-ttm on its own shows the current limit; amd-ttm --set 100 sets it to 100 GB after a reboot, and amd-ttm --clear restores the defaults. On the 64 GB version, choose a much lower value to leave memory for the system.
#What are the limitations of the GMKtec EVO-X2?
The GMKtec EVO-X2's main limitations are memory bandwidth, noise under load, non-upgradable memory, and the lack of CUDA, according to the official product page and reviews from ServeTheHome and Tom's Hardware.
- Noise and power consumption
- ServeTheHome measured 8 to 14 W and 37 to 38 dBA at idle, then 147 to 160 W and 41 to 43 dBA during Llama 3.3 70B Q6_K inference.
- Finition
- In its February 4, 2026 review, Tom's Hardware noted a cheap-looking case, an external power brick, a hard-to-access interior, and an audible fan.
- No CUDA
- The GPU is a Radeon: tools that require CUDA don't work; llama.cpp, LM Studio and Ollama run through ROCm or Vulkan.
Privacy depends on the software, not the machine: LM Studio, llama.cpp, and Ollama run the model locally. The AIPC app preinstalled by GMKtec was not examined for this guide; for sensitive data, use software you install and control yourself.
#GMKtec EVO-X2, EVO-X3, or EVO-X5 Pro: Buy Now or Wait?
Buy the GMKtec EVO-X2 128 GB now if your models fit in 96 GB and you want a local assistant today. Choose something else if you're targeting models larger than 96 GB, need CUDA, or want to connect an external graphics card.
| Your needs | Option | Why |
|---|---|---|
| MoE models up to 70 GB, tight budget | EVO-X2 128 GB + 2 TB, €3,499.99 | Up to 96 GB for the GPU, €100 less than the EVO-X3 |
| Same needs, with an external GPU later | EVO-X3 128 GB + 2 TB, €3,599.99 | Same Ryzen AI Max+ 395, OCuLink PCIe Gen4 x4 port for an external GPU |
| Models larger than 96 GB | EVO-X5 Pro 192 GB + 2 TB, €6,599.00 | Ryzen AI Max+ PRO 495 and 192 GB, at nearly double the price |
| Tools that require CUDA | NVIDIA machine (DGX Spark or GeForce card) | The Radeon 8060S doesn't run CUDA |
| Models under 30 GB, speed above all | PC with a 24 to 32 GB graphics card | Dedicated graphics memory, faster for models that fit in it |
Can the GMKtec EVO-X2 run a 70B model?+
Can you add memory to the GMKtec EVO-X2 later?+
Is the GMKtec EVO-X2 a good choice for coding with a local model?+
Is the GMKtec EVO-X2 noisy?+
Should you choose a NVIDIA DGX Spark over the GMKtec EVO-X2?+
#Sources and verification as of November 18, 2026
Specs and prices were collected on November 18, 2026 from GMKtec's official stores, settings from AMD and Ollama documentation, and measurements from ServeTheHome and Tom's Hardware reviews; prices, drivers, and required kernel versions may change.
- Source: official GMKtec EVO-X2 product page (gmktec.fr)
- Source: AMD, Ryzen AI Max+ 395 and Variable Graphics Memory
- Source: AMD, models up to 128 billion parameters on Windows
- Source: AMD ROCm documentation, RDNA 3.5 APU optimization
- Source: Ollama documentation, supported hardware
- Source: ServeTheHome, GMKtec EVO-X2 review
- Source: Tom's Hardware, GMKtec EVO-X2 review
- Source: NVIDIA, DGX Spark specifications
- Source: official GMKtec EVO-X3 product page
Feedback, an error, or a clarification? Let us know—it improves the guide for everyone.