iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Set up the GMKtec EVO-X1 Pro by updating Windows and its AMD/GMKtec drivers, checking the BIOS, then choosing a runtime that matches your workflow: GMKtec Claw for the quickest start, Ollama for a local API and automation, or LM Studio for graphical model discovery. Its 64GB of soldered LPDDR5-6400 memory is shared by Windows, the Radeon 890M integrated GPU and your model, so begin with a quantized 7B–14B model and measure before attempting larger ones. Add a PCIe 4.0 NVMe drive for a growing model library. An OCuLink PCIe 4.0 x4 port can support an external GPU, but it must only be connected or disconnected while the mini PC is fully powered off.
What does the GMKtec EVO-X1 Pro provide for local AI?
The EVO-X1 Pro is a Windows 11 Pro mini PC built around AMD’s Ryzen AI 9 HX 470. GMKtec specifies 12 CPU cores and 24 threads, Radeon 890M integrated graphics, up to 55 TOPS of dedicated NPU performance and up to 86 TOPS of total AI performance. Those TOPS figures are manufacturer specifications, not independent measurements of token generation speed. The practical limit for a local language model is usually available system memory, model quantization, context length and the software backend that can use the CPU, GPU or NPU.
| Component | GMKtec specification | What it means for local models |
|---|---|---|
| Processor | AMD Ryzen AI 9 HX 470, 12 cores / 24 threads | Plenty of CPU threads for inference, preprocessing and multitasking; actual speed varies by model and runtime. |
| Graphics | Radeon 890M integrated graphics | Uses the same system memory as Windows; GPU offload depends on runtime and supported backend. |
| NPU | Up to 55 TOPS dedicated | Useful only when an application has a compatible NPU path. Do not assume every LLM will use it. |
| Total AI figure | Up to 86 TOPS | A combined manufacturer rating, not a published LLM benchmark. |
| Memory | 64GB onboard LPDDR5-6400 | Fast shared memory, but it is soldered and cannot be upgraded. Reserve headroom for Windows and the runtime. |
| Storage expansion | Two M.2 2280 PCIe 4.0 x4 slots; PCIe 3.0 and 4.0 NVMe drives supported | Install a second NVMe drive for model files. SATA M.2 SSDs are not supported. GMKtec lists up to 16TB total capacity. |
| Power | 120W adapter included | Use the supplied adapter for sustained inference and firmware work. |
| External graphics | OCuLink PCIe 4.0 x4 | Can connect a desktop-class external GPU, but the connection is not hot-pluggable. |
At launch, independent performance data was not available. TechRadar’s July 29, 2026 launch coverage cautioned that the machine was brand new and had no independent reviews verifying GMKtec’s performance or AI TOPS claims. Treat the table above as the manufacturer’s configuration, then use your own measurements to decide which model and backend are worthwhile.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Which local-AI runtime should you use first?
Choose GMKtec Claw when you want the least setup, Ollama when another application will call a local API, and LM Studio when you prefer browsing and chatting with models in a graphical interface. You can install more than one; model files and settings are managed separately unless you deliberately share or relocate them.
#1 Best Overall
- Read Before You Buy — No Video Output: These adapters support charging and USB 2.0 data transfer, but cannot transmit video signals. Except for standard USB webcams (which use USB data only), they are not compatible with HDMI/DisplayPort cables, video-capable USB-C hubs, or docking stations with video output.
- Convert USB-A Ports to USB-C: Designed to connect USB-C earphones, cables, flash drives, card readers, and other USB-C accessories to standard USB-A ports. Plug-and-play with no drivers or software required.
- Aluminum Alloy Housing: Built with a sturdy aluminum alloy shell that aids in heat dissipation and protects against daily wear and scratches. Designed to maintain a stable and secure connection.
- Compact & Travel-Friendly: The ultra-compact design allows the adapter to stay plugged into your device without blocking adjacent ports or adding bulk, reducing wear and tear on your original USB ports.
- 12-Month Warranty: Backed by a 12-month manufacturer warranty for peace of mind. Designed to meet strict quality control standards for reliable everyday performance.
| Runtime | Setup effort | Interface and workflow | API or automation | Quantization and model workflow | GPU/NPU considerations |
|---|---|---|---|---|---|
| GMKtec Claw | Lowest because it is pre-installed | Start from the included application and its available model controls | Use the capabilities exposed by the installed version; verify its local endpoint before scripting | Use the models and quantization choices offered in the app | Backend support is tied to the shipped build; check its settings rather than assuming NPU use |
| Ollama | Low; one PowerShell command, then a model pull | Command line and local service, convenient for repeatable workflows | Strong fit for local HTTP API calls, scripts and application integrations | Pull model variants by name and test a small one before larger files | Ollama advises checking GPU and memory suitability; acceleration depends on the installed build and hardware support |
| LM Studio | Low; install the Windows build and use Discover | Graphical model browser, download manager and chat interface | Useful for interactive use; enable its local server when an application needs one | Shows quantized files and labels such as Q3 or Q8; choose higher fidelity when memory allows | Choose the backend exposed by the current release and confirm whether GPU offload is active |
| AMD ROCm path | Advanced | Developer-oriented installation and configuration | Best when a specific framework or application requires ROCm | Model handling is supplied by the framework you install | Check AMD’s current compatibility matrix for this processor, Radeon 890M, Windows version and framework before investing time |
How should you prepare Windows before loading a model?
Finish operating-system and firmware housekeeping before downloading multi-gigabyte model files. A stable, updated system makes later troubleshooting much easier.
- Connect the supplied hardware. Attach the 120W adapter, display, keyboard, mouse and network connection. Keep the mini PC on reliable power throughout updates and any BIOS operation.
- Run Windows Update. Install all ordinary updates and restart until Windows reports that no further updates are pending. Record the Windows build in Settings > System > About so you can correlate later driver behavior.
- Install current drivers. Download the current GMKtec and AMD graphics/chipset packages from GMKtec’s support resources. Reboot after chipset and graphics updates, even if the installer does not insist.
- Create a restore point. Open Start, search for Create a restore point, select the system drive, choose Configure if protection is disabled, then create a point before changing firmware or graphics settings.
- Check free space. Keep Windows and applications on the system drive and leave room for updates, temporary files and swap. Plan the second M.2 slot for models if your library will grow.
How do you check and update the EVO-X1 Pro BIOS safely?
Check the BIOS version before installing a model runtime. GMKtec lists BIOS EVO_X1_PRO_103_260812 for a startup-freeze issue reported after Windows update KB5101684. Use the In-OS Flash method only when Windows starts normally, keep stable power connected, select the EVO-X1 PRO image specifically, and never interrupt the flash.
- Press Win+R, enter
msinfo32, and read BIOS Version/Date. Write the value down. - Open GMKtec’s support page and compare the listed EVO-X1 PRO BIOS revision and release notes with your value. Do not use an image for another EVO model.
- If the notice applies, download the BIOS package and read its included instructions. The In-OS Flash route is intended for a normally booting Windows installation.
- Close applications, connect the adapter directly, and temporarily avoid sleep or shutdown actions. Start the flash utility as instructed by GMKtec.
- Allow every automatic restart to finish. A blank display or several minutes of inactivity can be normal during firmware work; do not press the power button unless GMKtec’s recovery instructions explicitly require it.
- After Windows returns, check
msinfo32again and confirm the intended revision. If the machine will not start, stop repeated attempts and use GMKtec’s documented recovery or support process.
How do you start with GMKtec Claw?
Claw is the sensible first run because it is already installed. Open it from the Start menu, inspect its model list and backend settings, and load the smallest instruct model it offers. Send a short prompt, confirm that a response is generated, and note whether the interface reports CPU or GPU use. This baseline tells you that Windows, drivers and the included runtime are functioning before you add another variable.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problemsDo not infer NPU utilization from the EVO-X1 Pro’s 55-TOPS specification. If Claw exposes an NPU toggle, test it against the default path with the same prompt and context, and keep the faster or more stable configuration. Record the application version because a later update can change available backends.
How do you install Ollama on Windows?
Ollama is the best first choice when scripts, a local API or repeatable command-line workflows matter more than a visual model catalog.
- Open PowerShell from the Start menu.
- Run the official Windows installer command exactly as shown:
irm https://ollama.com/install.ps1 | iex. - Accept the installer prompts. Restart Windows if the installer requests it.
- Pull a small instruct model first, using the model name shown in Ollama’s current library. Avoid starting with the largest file available.
- Send a short test prompt and confirm that generation completes. Watch Task Manager for memory pressure and note whether the runtime reports GPU acceleration.
- Only after the baseline works, try a larger or higher-fidelity quantization. Ollama explicitly recommends checking whether your GPU and memory are suitable; speed is hardware- and model-dependent.
Keep the command-line session and model name in your notes. If a model fails to load, the exact tag, quantization and error text are more useful than a general statement that “Ollama is slow.”
Rank #2
- 5-in-1 USB-C Hub: Experience comprehensive connectivity featuring a Power Delivery input, two USB-A 2.0 ports, a USB-A 3.0 port, and an HDMI port. (Note: The USB-C power delivery input port is only for connecting an external wall charger to power your laptop and cannot power peripheral devices.)
- 90W Pass-Through Charging: Achieve optimal charging with 90W pass-through power to your laptop, supported by a total input of 100W, with the hub reserving 10W for operational efficiency. (Note: Wall charger not included.)
- Quick Data Transfers: Accelerate your productivity with rapid data transfers using a high-speed 5Gbps USB 3.0 port and two 480Mbps USB 2.0 ports.
- 4K HDMI Display: Enhance your visual experience with a hub capable of delivering 4K resolution at 30Hz in both mirror and extend modes. Please note that this hub is compatible with MacBook (macOS 12 and newer), Windows 10 and 11, ChromeOS, and laptops equipped with DP Alt Mode and Power Delivery. Note: This device is not compatible with Linux.
- What You Get: Anker USB-C Hub (5-in-1, 4K HDMI), welcome guide, 18-month warranty, and our friendly customer service.
How do you install and use LM Studio?
LM Studio is convenient when you want to discover models, compare quantized files and chat without memorizing commands.
- Install the current Windows build from LM Studio’s official distribution.
- Open Discover and search for an instruct model whose file size is appropriate for the available memory.
- Choose a quantized file. Labels such as Q3 and Q8 describe the quantization level; higher numbers generally preserve more weight precision but require more memory.
- Download one file, open it in the chat area, and keep the initial context length modest.
- Run a repeatable prompt, then inspect the app’s memory and GPU-offload indicators. Increase context or fidelity only when the baseline is stable.
- If another application needs a local endpoint, enable LM Studio’s local server and test it from that application before downloading additional models.
A model that appears in Discover is not automatically a good fit for this mini PC. File size, context allocation and the portion offloaded to the Radeon 890M determine whether Windows remains responsive.
When should you use AMD ROCm instead?
Use ROCm when a particular framework or application requires AMD’s acceleration stack and you are prepared to troubleshoot compatibility. It is not a prerequisite for ordinary Claw, Ollama or LM Studio use.
Before installing, check AMD’s current ROCm compatibility matrix for the exact operating system, Ryzen/Radeon device and framework version. Support can differ between Linux and Windows and can change between releases. Install only the components required by your target application, then run that application’s own verification command. If the matrix does not list your configuration, expect CPU fallback or unsupported behavior rather than assuming the 86-TOPS figure will apply.
What model size fits in 64GB of shared memory?
Start with a quantized 7B–14B model. The EVO-X1 Pro’s 64GB is shared by Windows, the integrated GPU, the runtime, the model weights and the context cache; it is not 64GB dedicated solely to an LLM. Larger models can sometimes load with aggressive quantization, but they should be treated as experiments until measured on your exact installation.
A useful planning estimate for weights alone is parameter count × bits per parameter ÷ 8. It excludes runtime overhead, tokenizer data, temporary buffers and the additional memory required by a longer context. For example, a 7B model at 4-bit precision is roughly 3.5GB of raw weights, while a 14B model at 4-bit is roughly 7GB; the process will consume more. Higher-fidelity formats increase that requirement, and longer context can become the limiting factor even when the weight file fits.
Rank #3
- Sleek 7-in-1 USB-C Hub: Features an HDMI port, two USB-A 3.0 ports, and a USB-C data port, each providing 5Gbps transfer speeds. It also includes a USB-C PD input port for charging up to 100W and dual SD and TF card slots, all in a compact design.
- Flawless 4K@60Hz Video with HDMI: Delivers exceptional clarity and smoothness with its 4K@60Hz HDMI port, making it ideal for high-definition presentations and entertainment. (Note: Only the HDMI port supports video projection; the USB-C port is for data transfer only.)
- Double Up on Efficiency: The two USB-A 3.0 ports and a USB-C port support a fast 5Gbps data rate, significantly boosting your transfer speeds and improving productivity.
- Fast and Reliable 85W Charging: Offers high-capacity, speedy charging for laptops up to 85W, so you spend less time tethered to an outlet and more time being productive.
- What You Get: Anker USB-C Hub (7-in-1), welcome guide, 18-month warranty, and our friendly customer service.
| Model class | Typical starting quantization | Raw-weight planning estimate | Recommended use on this machine | What to measure |
|---|---|---|---|---|
| 7B–8B | 4-bit or higher | Approximately 3.5–4GB at 4-bit, before overhead | Best baseline for responsive chat and troubleshooting | Prompt processing speed, generation speed, memory use and GPU offload |
| 12B–14B | 4-bit; try higher fidelity if headroom remains | Approximately 6–7GB at 4-bit, before overhead | Good next step when the smaller model is stable | Whether context growth causes paging or stutter |
| 20B–34B | Aggressive quantization may be required | Approximately 10–17GB at 4-bit, before overhead | Experimental; reserve substantial memory for Windows and runtime buffers | Load success, sustained tokens per second and responsiveness |
| 70B class | Usually requires very aggressive quantization on this hardware | Approximately 35GB at 4-bit, before overhead | Possible only as a carefully measured experiment; context and shared-memory overhead can make it impractical | Whether the model remains usable after a long prompt, not just whether it loads |
Quantization is a trade-off, not a speed guarantee. A Q3 file can fit where Q8 cannot, but may reduce answer quality; a higher-precision file can improve fidelity while increasing memory pressure. Compare files with the same prompt and context so that the result is meaningful.
How should you choose context length and quantization?
Set the context window to the smallest length that serves your task, then increase it deliberately. Every additional token in the context requires cache memory, and a model that is comfortable at 4,096 tokens may become sluggish at a much larger window even when its weight file is unchanged.
- Chat and short coding tasks: begin with the runtime’s modest default context and a 4-bit model.
- Long documents: increase context in stages and watch committed memory rather than relying on the model file size.
- Higher answer fidelity: test Q5, Q6 or Q8 only after confirming that the model and context fit without swapping.
- Repeatable comparisons: keep prompt text, context length, temperature and output limit constant while changing one variable.
How do you add storage for a growing model library?
Use the second M.2 2280 slot with a PCIe 3.0 or PCIe 4.0 NVMe SSD. Do not buy a SATA M.2 drive: GMKtec lists SATA SSDs as unsupported. The manufacturer lists up to 16TB total capacity across the supported slots.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →- Shut down Windows completely, unplug the adapter and hold the power button briefly to discharge residual power.
- Open the case according to GMKtec’s service instructions. Avoid touching exposed contacts and keep screws organized.
- Insert the NVMe module at a slight angle into the empty M.2 2280 connector, press it flat and secure it with the retaining screw.
- Reassemble the case, reconnect power and boot Windows.
- Open Disk Management, initialize the new disk using the partition style Windows recommends for a modern UEFI system, create a volume and assign a drive letter.
- Point your runtime’s model directory at the new volume where that application supports a custom location. Keep Windows, drivers and frequently used applications on the system drive.
Storage capacity does not increase usable model memory. A larger SSD lets you keep more model files; it does not make a model run faster or allow a model that exceeds available RAM to fit comfortably.
Do you need an eGPU through OCuLink?
You do not need an eGPU for initial 7B–14B testing. The integrated Radeon 890M and shared 64GB memory are the simplest starting point. An OCuLink PCIe 4.0 x4 connection becomes useful when you need more graphics memory or sustained throughput for larger models and can accept the extra cost, power draw, desk space and noise of an external GPU enclosure and its power supply.
OCuLink is not a hot-plug connector on this system. GMKtec states: “Please connect or disconnect OCuLink devices only when the Mini PC is fully powered off.” Shut down Windows, wait until the machine is off, then attach or remove the dock or GPU. Power the system back on only after the connection is secure. Reverse the process before disconnecting it. Install the external GPU’s current driver, verify that the runtime detects it, and retest your model; an eGPU does not automatically make every backend use the card.
Rank #4
- Dual Converters, Infinite Potential：Includes 2× USB C male to USB A female adapters and 2× USB A male to USB C female adapters. Perfect for a wide range of uses—tablets with Bluetooth keyboards, expand USB ports on macbook, and more. Two different converters for all your daily needs
- Next-Level 10Gbps & 3A Charging: No more slow 480Mbps, this usb to usb c adapter has a transfer speed of up to 10Gbps, allowing you to do more transferring in less time. This usb adapter fits both USB A and USB C charger, supporting up to 3A fast charging
- Upgraded Exquisite Craftsmanship: With an aluminum alloy housing and metal connector, the usbc to usb adapter is extremely durable and sturdy. Rigorously tested to withstand more than 10,000 times of plugging and unplugging, ensuring long-lasting performance
- Broad Compatible: The usb c to usb adapter widely supports all USB C/ USB A devices like laptops, tablets, cellphones, car chargers, and phone chargers. Such as compatible with MacBook Pro/Air 2023/2022, Thunderbolt 4/3 Devices,Apple MagSafe Watch 9/8/7/SE/Ultra, iPad Pro 2022/2021, Samsung Galaxy S23/S20/S10, and iPhone 17/16/15 Pro. Plug and play
- Please Note: To reach 10Gbps speed, keep the cable under 3.3 ft. For USB A Male to USB C adapters, try flipping the USB C connector. USB C Male to USB A adapters support bidirectional 10Gbps transfer within 3.3 ft
How do you measure whether a model is actually usable?
Record each test as a small experiment instead of relying on a single impression. Published TOPS ratings do not tell you the tokens-per-second result for your model, quantization and context.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →| Record | Why it matters |
|---|---|
| Runtime and version | Backend behavior and acceleration support can change after updates. |
| Model name and exact file/tag | Two files with the same parameter count can have different quantization and architecture. |
| Quantization | Explains fidelity and memory differences between otherwise similar files. |
| Context length and output limit | Controls cache allocation and makes runs comparable. |
| Prompt-processing speed | Shows how quickly a long input is ingested. |
| Generation speed | Shows sustained response speed after the prompt is processed. |
| Peak memory and GPU allocation | Reveals whether Windows or the runtime is close to paging. |
| GPU-offload status | Confirms whether acceleration is active instead of assumed. |
| Thermal and power behavior | Long sessions can throttle even when a short test looks fast. |
- Close unrelated applications and let the system idle for a minute.
- Use the same prompt, context and output limit for every candidate.
- Run at least one short warm-up generation, then record the next run.
- Repeat after a longer prompt or a sustained conversation to expose cache and thermal limits.
- Keep the result as a user measurement for your configuration, not as an EVO-X1 Pro benchmark that applies to every unit.
What problems are most likely during setup?
Windows freezes at startup after an update
Check whether the incident follows Windows update KB5101684 and compare your BIOS with GMKtec’s listed EVO_X1_PRO_103_260812 fix. If Windows still starts, use the documented In-OS Flash process with stable power. If it does not, use GMKtec’s recovery instructions rather than repeatedly forcing power cycles.
The model downloads but will not load
Reduce the quantization level or context length, close other memory-heavy applications and retry with a 7B–8B model. Check committed memory in Task Manager; a file that fits on the SSD can still exceed available shared memory once runtime buffers are allocated.
Generation is unexpectedly slow
Confirm that the runtime reports the intended GPU backend and that the graphics driver is current. Compare a shorter context and a smaller quantized model. If GPU offload is unavailable, CPU inference can still work, but its speed should be recorded rather than inferred from the TOPS specification.
The second SSD is missing
Verify that it is an NVMe PCIe 3.0 or 4.0 drive seated in the M.2 2280 connector. A SATA M.2 device is unsupported. If the NVMe drive is absent from Disk Management and BIOS, power down, reseat it and check the drive’s own documentation before changing firmware.
Free tools Windows power users keep installed
One-click scans. No signup required.
An OCuLink GPU is not detected
Power the mini PC fully off, reseat the cable and dock, then boot again. Install the external GPU driver and check the runtime’s backend selection. Do not connect or remove the OCuLink device while the computer is running.
Best Value
- 5-in-1 Connectivity: Equipped with a 4K HDMI port, a 5 Gbps USB-C data port, two 5 Gbps USB-A ports, and a USB C 100W PD-IN port. Note: The USB C 100W PD-IN port supports only charging and does not support data transfer devices such as headphones or speakers.
- Powerful Pass-Through Charging: Supports up to 85W pass-through charging so you can power up your laptop while you use the hub. Note: Pass-through charging requires a charger (not included). Note: To achieve full power for iPad, we recommend using a 45W wall charger.
- Transfer Files in Seconds: Move files to and from your laptop at speeds of up to 5 Gbps via the USB-C and USB-A data ports. Note: The USB C 5Gbps Data port does not support video output.
- HD Display: Connect to the HDMI port to stream or mirror content to an external monitor in resolutions of up to 4K@30Hz. Note: The USB-C ports do not support video output.
- What You Get: Anker 332 USB-C Hub (5-in-1), welcome guide, our worry-free 18-month warranty, and friendly customer service.
Which setup is the best starting point?
Use Claw to prove that the freshly updated EVO-X1 Pro can generate locally, then move to Ollama if you need scripts or an API, or LM Studio if you want visual model discovery. Keep your first model in the 7B–14B quantized range, add a PCIe 4.0 NVMe SSD when storage becomes the constraint, and consider OCuLink only when integrated graphics and shared memory no longer meet a measured workload. Recheck the current runtime and ROCm compatibility documentation whenever you update Windows, drivers or the applications.
Frequently Asked Questions
How can I keep a long generation from being interrupted by Windows sleep?
Open Settings > System > Power & battery > Screen and sleep and set the plugged-in sleep timeout to Never for the duration of the run. Restore a normal timeout afterward so the mini PC does not remain awake unnecessarily.
Can Ollama store its model library on the second NVMe drive?
Yes, on Windows you can set the OLLAMA_MODELS environment variable to a folder on the new drive, then restart the Ollama service. Move or re-pull models after changing the location and verify the service can read the directory.
Why does a model use more memory than the download size shown in the browser?
The download is mainly the weight file. The runtime also allocates tokenizer data, temporary buffers and a key-value cache for the context, while Windows and the integrated GPU share the same 64GB pool.
Is the NPU automatically used by every local LLM application?
No. An application must provide a compatible NPU execution path, and support varies by model format, framework and Windows build. Check the runtime’s backend or device indicator instead of treating the advertised TOPS number as proof of NPU usage.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
