Gmktec Evo-X3 Mini-PC Review: Looking at a 128GB AMD Local AI Box

Welcome back to Self-Hosted Tech. If you’ve been hanging around the homelab and self-hosting community for a while, you already know that RAM is king. But when you start dipping your toes into the deep end of the pool with local AI, RAM becomes the emperor, the grand vizier, and the entire royal court all at once.

We’ve seen a massive influx of mini-PCs hitting the market lately, but almost all of them cap out at 32GB or maybe 64GB of RAM. That’s fine for running a few Docker containers or a small Proxmox cluster, but it’s a massive bottleneck for running large language models (LLMs) locally. Enter the Gmktec Evo-X3.

This little box looks like your standard mini-PC, but it’s packing a ridiculous 128GB of DDR5 RAM alongside an AMD Ryzen processor. Today, we’re going to look at whether this 128GB AMD local AI box is the ultimate budget server for homelabbers and AI tinkerers, or if it’s just a gimmick. Let’s dive in.

The Specs: What’s Under the Hood?

Before we get into the AI side of things, let’s look at what the Gmktec Evo-X3 is actually bringing to the table.

  • CPU: AMD Ryzen 7 8845HS (8 Cores, 16 Threads, up to 5.1GHz)
  • RAM: 128GB DDR5-5600MHz (Dual Channel, 2x 64GB SODIMMs)
  • Storage: Dual M.2 2280 PCIe 4.0 NVMe slots
  • Network: 2x 2.5GbE LAN ports
  • GPU: Integrated AMD Radeon 780M (RDNA 3 architecture)
  • Ports: USB4, multiple USB 3.2 ports, HDMI, DisplayPort
  • OS Support: Windows 11 Pro, Ubuntu, Debian, Proxmox

On paper, this is a powerhouse. The Ryzen 7 8845HS is an incredibly efficient mobile chip that sips power but offers fantastic multi-threaded performance. The inclusion of dual 2.5GbE ports is a massive win for homelabbers who want to run OPNsense or pfSense, or just want link aggregation for their NAS setups.

But the star of the show is undeniably that 128GB of DDR5 RAM.

Design and Build Quality

The Evo-X3 doesn’t try to reinvent the wheel when it comes to chassis design. It’s a standard 1-liter mini-PC box made of a mix of plastic and metal. It’s VESA mountable, meaning you can hide it behind a monitor or tuck it away in a network rack.

Opening it up is straightforward. You pop off the bottom panel, and you’re greeted by the dual NVMe slots, the Wi-Fi module, and the two RAM slots. Because 64GB SODIMMs are physically dense, Gmktec has done a decent job with thermal padding to ensure the RAM doesn’t overheat during sustained workloads.

Cooling is handled by a blower-style fan. Under normal homelab loads (running Docker containers, Pi-hole, Home Assistant), the unit is virtually silent. However, when you start loading up a 70-billion parameter AI model, the fan will spin up. It’s not jet-engine loud, but you’ll definitely know it’s working.

The Elephant in the Room: AMD vs. Nvidia for AI

Let’s address the elephant in the room. If you want to do local AI, Nvidia is the easy button. CUDA is the industry standard, and almost every AI tool out there supports it natively out of the box. AMD’s alternative, ROCm, has gotten significantly better over the last year, but it can still be finicky to set up on Linux.

So, why would you buy an AMD mini-PC with an integrated GPU for AI?

The answer lies in the 128GB of system RAM.

When you run an AI model using tools like Ollama or llama.cpp, the software will try to load the model into VRAM (Video RAM) for maximum speed. If the model is too big for your VRAM, the software offloads the rest of the model into your standard system RAM and processes it using your CPU.

A dedicated RTX 4090 has 24GB of VRAM and costs over $1,500. The Gmktec Evo-X3 has 128GB of unified memory (shared between the system and the integrated Radeon 780M) and costs a fraction of that price.

Is it as fast as an RTX 4090? Not even close. DDR5 system memory has a bandwidth of around 80-90 GB/s, while an RTX 4090 has a bandwidth of over 1,000 GB/s. This means the Evo-X3 will generate text slower. But—and this is the crucial part for homelabbers—it will actually run models that are physically impossible to run on consumer Nvidia GPUs.

Practical Walkthrough: Running a Massive Local LLM

Let’s get practical. I wiped the included Windows installation and installed Ubuntu Server 22.04 LTS. If you want to follow along, the process is identical on Debian or even inside a Proxmox Linux Container (LXC).

We are going to use Ollama, which is the absolute best beginner-friendly tool for running local AI. It uses llama.cpp under the hood, which handles the CPU/RAM offloading automatically.

Step 1: Install Docker and Ollama

First, let’s make sure our system is up to date and install Docker, because everything in a homelab should run in Docker.

1
2

sudo apt update && sudo apt upgrade -y
sudo apt install docker.io docker-compose -y

Next, we’ll run Ollama in a Docker container. We need to pass the /dev/kfd and /dev/dri devices to the container so it can access the AMD integrated GPU, though for this massive model, we’ll be leaning heavily on the system RAM anyway.

1
2
3
4
5
6

docker run -d --name ollama \
  -p 11434:11434 \
  -v ollama_data:/root/.ollama \
  --device /dev/kfd \
  --device /dev/dri \
  ollama/ollama:latest

Step 2: Pulling a Massive Model

With Ollama running, let’s test the limits of that 128GB RAM. We are going to pull a heavily quantized version of Llama 3 70B. Quantization compresses the model, reducing its size and accuracy slightly, but making it possible to run on consumer hardware.

A 4-bit quantized Llama 3 70B model takes up roughly 40GB of space. On a 32GB system, this would crash or fail to load. On the Evo-X3, it fits comfortably with plenty of room for the OS and your other homelab services.

Open a shell into your Ollama container:

1

docker exec -it ollama bash

Now, pull and run the model:

1

ollama run llama3:70b-instruct-q4_K_M

You’ll watch the download progress, and then you’ll see the model load into memory. You can open a second SSH session to your Ubuntu server and run htop to watch the magic happen. You’ll see your RAM usage jump by about 40-45GB.

Step 3: Generating Text

Once the model is loaded, you’ll be presented with a prompt. Try asking it something complex, like writing a Python script or explaining a complex networking concept.

1

>>> Write a bash script to backup my docker volumes to a remote server using rsync.

You will notice the text generation speed. Because it’s running largely on system RAM via the CPU, you might get around 3 to 5 tokens per second. It’s not lightning fast, but it is highly usable for background tasks, coding assistance, or document summarization.

The true power here is that you can leave this running 24/7 in your homelab. You can hook it up to tools like Open WebUI, Dify, or AnythingLLM, giving your whole household access to a private, uncensored, massive AI model without paying OpenAI API fees.

Beyond AI: The Ultimate Proxmox Homelab Node

While the AI capabilities are the flashy headline, let’s not forget what Gmktec has accidentally built here: an absolute monster of a virtualization node.

128GB of RAM in a 1-liter box that draws less than 50 watts under typical load is a homelaber’s dream. If you install Proxmox VE on the Evo-X3, the possibilities are endless. Here is just a sample of what you can run simultaneously on this single box:

  • TrueNAS Scale VM: Allocate 64GB of RAM to a TrueNAS virtual machine. ZFS loves RAM for its ARC (Adaptive Replacement Cache). With 64GB, your ZFS pool will cache an enormous amount of your most frequently accessed files, making your network shares incredibly fast.
  • Docker Host LXC: Allocate 16GB to a dedicated LXC container running Portainer and all your standard self-hosted apps (Pi-hole, Home Assistant, Jellyfin, Vaultwarden).
  • Kubernetes Cluster: Spin up three lightweight VMs with 8GB of RAM each to create a local K3s cluster for testing cloud-native deployments.
  • Windows VM: Allocate 16GB to a Windows VM for remote desktop access when you need to run Windows-only software from your Mac or Linux laptop.

Because the Ryzen 7 8845HS has 8 cores and 16 threads, you have enough compute power to juggle all of these workloads without breaking a sweat. And with dual 2.5GbE ports, you can dedicate one port to your storage network and one to your management/Docker network, keeping traffic isolated and fast.

Pros and Cons

After a few weeks of testing the Evo-X3, here is my final breakdown:

Pros:

  • Massive RAM Capacity: 128GB in a mini-PC is incredibly rare and opens up use cases (like massive LLMs and heavy ZFS caching) that are usually reserved for large tower servers.
  • Power Efficiency: The AMD Ryzen 8845HS is incredibly power-efficient. You can run this 24/7 and barely notice a bump in your electric bill.
  • Dual 2.5GbE: Perfect for homelab networking, router setups, and link aggregation.
  • Dual NVMe Slots: Allows for fast storage redundancy or separating your OS from your data drives.

Cons:

  • AI Speed: Running models on system RAM is slower than dedicated VRAM. If you need real-time, high-speed AI generation, you still need an Nvidia GPU.
  • AMD ROCm Friction: While Ollama makes it easy, trying to get other, less-polished AI tools running on the AMD iGPU can require compiling code and tweaking Linux drivers.
  • Thermals: When running a 70B AI model, the box gets hot and the fan gets loud. It’s a small chassis doing heavy lifting.

Summary

The Gmktec Evo-X3 is a fascinating piece of hardware. It sits in a weird, wonderful niche where mini-PC, homelab server, and budget AI workstation all intersect.

If your primary goal is to run Stable Diffusion at lightning speed or get 50 tokens per second on ChatGPT-sized models, this isn’t the box for you—you need a dedicated desktop with an Nvidia GPU. But if you are a self-hoster who wants a silent, low-power box that can run your entire Docker stack, act as a ZFS storage beast, and also run massive 70B+ local AI models in the background, the Evo-X3 is a phenomenal option. The 128GB of RAM fundamentally changes what a mini-PC is capable of doing.

What do you think? Are you ready to ditch the giant tower servers for a 1-liter box packed with RAM? Let us know in the comments below what model you’d run first on 128GB of RAM.

If you found this review helpful, make sure to hit that subscribe button on the Self-Hosted Tech YouTube channel and drop your email on the blog for more practical guides on homelabs, local AI, and self-hosting. Until next time, keep your containers running and your data local!