GMKtec EVO-X2 for local LLM and Ollama: is Strix Halo worth it?

Short answer: the GMKtec EVO-X2 is a Strix Halo mini PC built around the AMD Ryzen AI Max+ 395 with soldered LPDDR5X-8000 memory in 64GB, 96GB or 128GB configs. Buy it when you need capacity for 70B-class and large MoE models in one quiet-ish box. Skip it if you only run 8B to 14B models, or if you want the fastest tokens per second (a desktop with a discrete GPU still wins on speed).

Every hardware claim below is from AMD's Ryzen AI Max+ 395 product page or GMKtec's EVO-X2 product page, rechecked Nov 6, 2026. No prices here. Click through for the current price.

What is inside the EVO-X2

RAM configs: buy capacity up front

Memory is soldered. There are no SO-DIMM slots. That is the single most important owner decision on this box.

• 64GB: fine for 14B to 32B daily drivers and light multitasking. GMKtec's own LLM table marks DeepSeek-R1 70B and Llama 4 109B as not supported on the 64GB SKU.
• 96GB: middle SKU GMKtec lists in its compatibility table. Useful if you want more headroom than 64GB without jumping to 128GB.
• 128GB: the config that makes Strix Halo make sense for local AI. GMKtec lists DeepSeek-R1 70B, Llama 4 109B and Qwen3 235B as supported on the 128GB model. Up to 96GB can be carved out for the iGPU after a BIOS update on some units.

If you are still deciding between 32GB upgradeable mini PCs and this box, start with our best mini PC for local AI comparison and how much RAM and VRAM Ollama needs.

Bandwidth vs a discrete GPU desktop

Token generation is mostly a memory-bandwidth problem once the model fits. AMD's 256-bit LPDDR5X-8000 pool peaks near 256 GB/s. That is roughly double a typical thin APU, but far below a modern discrete card (often 400 to 1,000+ GB/s depending on the card).

What that means in practice:

• Capacity win: a 43GB 4-bit Llama 3.3 70B file (Ollama's default) fits in the 128GB pool with room for context. A single RTX 3090 or 4090 (24GB) or even an RTX 5090 (32GB) cannot hold it fully on the GPU. See how much VRAM Llama 3.3 70B needs.
• Speed loss: dense 70B decode on ~256 GB/s will be much slower than the same model fully on a high-bandwidth NVIDIA card or a multi-GPU split. Buy the EVO-X2 for "it fits and runs privately," not for "it is the fastest tok/s."
• MoE models help: sparse models only stream active experts per token, so they feel faster than a dense model of the same total size. GMKtec publishes its own token/s figures for GPT-OSS 120B and Qwen3 235B on the 128GB SKU; treat those as manufacturer numbers, not independent lab results.

iGPU expectations (Radeon 8060S)

The Radeon 8060S is a strong integrated GPU (40 CUs), not a substitute for CUDA. On Windows and Linux you will typically use Vulkan or AMD's ROCm stack with Ollama or llama.cpp. After you load a model, run ollama ps and check the PROCESSOR column. If it shows mostly CPU, you still have driver or backend work to do.

Compared with a tower:

• Choose EVO-X2 when you want one small always-on box that can hold big models without a second GPU, dual PSU math, or a loud mid-tower.
• Choose a discrete GPU desktop when 8B to 32B models at high tokens/s matter more than fitting 70B+. Our best budget GPU for Ollama guide covers the 12GB to 24GB cards that win on speed per dollar.

Thermals, noise and desk life

GMKtec rates the box at 120W sustained and 140W peak, cooled with heat pipes and dual fans, powered by a ~230W external brick. That is a lot of heat for a ~2.7 liter chassis. Owner reviews on GMKtec's own product page and third-party reviews often mention:

• Fan noise under heavy CPU load (more than under pure GPU load for some owners)
• BIOS updates needed for fan curves and to unlock higher VRAM allocation (up to 96GB on the 128GB SKU)
• Better cooling when the unit stands vertically so intake vents stay clear

Our guidance: treat it as a compact workstation, not a silent Mac mini. If silence is the top priority, compare with a Mac mini M6 / M5 Pro config instead.

OS and software path

GMKtec ships or supports Windows 11 Pro and lists Ubuntu. For a Windows path, follow our Ollama + Open WebUI install guide. On Linux, many Strix Halo owners prefer Ubuntu 24.04 with current AMD GPU packages. Either way:

1. Update BIOS / EC firmware first (GMKtec posts updates on its support channels).
2. Set the GPU memory / VRAM allocation in BIOS if you need more than the default share.
3. Install Ollama, pull a model that fits (start with an 8B or 14B), then run ollama ps.
4. Only then pull a 70B or large MoE file. Those downloads are tens of GB.

Who should buy it, who should skip it

• Buy the 128GB EVO-X2 if: you want 70B or large MoE models local, prefer one small box over a dual-GPU tower, and can live with fan noise under load.
• Buy 64GB only if: you like the chassis and ports but are sure you will stay at 32B and below. Otherwise save money with a SER8-class mini PC and upgradeable DDR5 (see the mini PC guide).
• Skip it if: you only need 8B chat, you want maximum tok/s on models that already fit in 12 to 24GB VRAM, or you need a silent bedroom PC.
• Also consider: a used dual RTX 3090 tower for 48GB of real VRAM when speed on dense 70B matters more than size and power draw. Math for that path is in the 70B VRAM guide and the local vs cloud cost calculator.

Running AI for a small office from one box? See private AI for small business. New to the whole idea? Start at the Local AI hub.

Sources (checked Nov 6, 2026): AMD Ryzen AI Max+ 395 product page (cores, clocks, cTDP 45-120W, 256-bit LPDDR5x-8000, max 128GB, Radeon 8060S 40 CU / 2900 MHz, NPU 50 TOPS / 126 TOPS overall). GMKtec EVO-X2 product page (120W sustained / 140W peak, ~230W brick, ports, dual M.2, Wi-Fi 7, 2.5GbE, OS, LLM support table by SKU, manufacturer token/s figures). Bandwidth 256 GB/s is our calculation from AMD's memory width and speed. "Who should buy / skip" and noise guidance are ours, based on those specs plus public owner comments on GMKtec's page. Not paid by GMKtec or AMD.

FAQ

Is the GMKtec EVO-X2 good for Ollama?

Yes, if you need memory capacity more than raw token speed. The 128GB LPDDR5X config can hold 70B-class models that will not fit on a 24GB or 32GB discrete GPU. Bandwidth is about 256GB/s theoretical, so dense 70B replies are slower than on a dual-GPU desktop.

How much RAM should I buy on the EVO-X2?

Buy the 128GB config if local 70B or large MoE models are why you want the box. Memory is soldered LPDDR5X and cannot be upgraded later. The 64GB config is fine for 14B to 32B daily drivers, but it is not the reason to pick Strix Halo over a cheaper mini PC.

Does the Radeon 8060S beat an RTX 3060 or 3090 for local AI?

Not on speed. A discrete NVIDIA card has much higher memory bandwidth (hundreds of GB/s more) and mature CUDA support. The EVO-X2 wins on capacity: up to 96GB of the 128GB pool can be allocated to the iGPU, which no consumer 24GB or 32GB card can match.

Is the EVO-X2 quiet?

It depends on load. GMKtec rates sustained power at 120W with peaks to 140W and ships a ~230W brick. Owner feedback and reviews often call out fan noise under heavy CPU load. Plan for a desk or shelf where noise is acceptable, not a silent bedroom setup.

Windows or Linux on the EVO-X2?

GMKtec lists Windows 11 Pro and Ubuntu. Ollama runs on both. On AMD you are in the ROCm or Vulkan world, not CUDA, so check ollama ps after you load a model to confirm what is running on the GPU.