GMKTEC EVO-X2 RUNS A 235B MODEL. SAME TIER AS THE TOP OPENAI AND CLAUDE PLANS

I found out about this late. Don't make the same mistake.

Follow & Bookmark this - I'm @starmexxx, I track how AI tools are creating new income streams most people haven't heard of yet. This one surprised even me.

At CES 2026 in Las Vegas, AMD CEO Lisa Su stood on stage for her keynote with a small black box sitting on the platform behind her. Not a server rack. Not a data center render. A mini PC the size of a hardcover book.

A few months later, at AMD's AI Developer Day in Shanghai, she walked up to that same device and personally signed it. The device is the GMKtec EVO-X2. It costs $1,700-2,000 depending on configuration. And it runs AI models so large that most cloud subscriptions can't touch them.

1/

The chip that changed the math.

Everything starts with AMD's Ryzen AI Max+ 395, codenamed Strix Halo. This is the first x86 chip ever built that can run a 200 billion parameter model on a single piece of silicon, using a unified memory architecture similar to Apple Silicon.

What that means in plain terms: instead of a separate graphics card with its own limited memory, the CPU and GPU share one giant pool of memory, up to 128GB of it. The model loads once and both the processor and graphics chip read from the same place.

2/

What the EVO-X2 actually does.

In real testing, AMD claimed the chip outperformed an NVIDIA RTX 5080 by more than 3x on DeepSeek R1 inference. That's a chip built into a mini PC the size of a lunchbox beating a $1,000+ discrete graphics card on a real AI workload.

With the 128GB configuration, up to 110GB can be allocated as usable VRAM on Linux. That's enough to run Qwen3-235B fully and smoothly, plus models like DeepSeek-V3 and Llama 3.3 70B without any quantization tricks.

For comparison, running a model that size through an API would put you firmly in $200/month territory, and even those plans often rate-limit you during peak hours.

3/

The subscriptions this replaces.

The EVO-X2 costs $1,700-2,000 once. Running it 24/7 at typical inference loads costs roughly $9 a month in electricity. Break-even against a $200/month subscription habit lands around 9-10 months. After that, every month is pure savings - over three years that's roughly $13,000 back in your pocket.

4/

Setting it up takes about ten minutes.

The EVO-X2 ships with Windows 11, but most people running serious local AI workloads put it on Linux to unlock the full 110GB of usable VRAM instead of Windows' 96GB cap.

Once the OS is set up, installing Ollama is a single command in the terminal. After that, pulling a model is one more command, and within fifteen to twenty minutes you have a fully private AI running locally with an interface that looks and feels exactly like ChatGPT through Open WebUI.

For developers, Claude Code can be pointed at the local model with a single environment variable change. Same interface, same commands, same workflow, except nothing leaves the machine and nothing costs money per request.

5/

The honest tradeoffs.

This isn't a Mac Mini. It's bigger, runs hotter under load, and pulls more electricity - about $9/month versus $2-3 for the Mac Mini setup. It's also a Windows-first device, though Linux support for AI workloads is well documented and widely used already.

The price is roughly 3x a base Mac Mini, but the capability difference is real - this thing runs models 3-4x larger than what fits comfortably in a Mac Mini's memory. If your AI usage genuinely needs 70B+ parameter models running locally, this is the tier where that becomes possible without spending $4,000+ on a Mac Studio.

Where cloud models still have an edge is in the most demanding agentic coding across massive codebases, and in absolute frontier-level reasoning where every percentage point of benchmark performance matters. For everything else, the gap has essentially closed.

The window.

AMD built a chip specifically so that running a 200 billion parameter model locally on consumer hardware would stop sounding like science fiction. GMKtec built that chip into a $1,700 box and AMD's own CEO put it on stage and signed it with her name.

The subscriptions made sense when nothing on a desk could touch what the cloud offered. For models in the 70B-235B range, that gap is basically gone now if you're willing to make a one-time hardware purchase instead of a recurring one.

A single $1,700 purchase, roughly $9 a month in electricity, and a 235 billion parameter model running quietly on your desk.

// The window is open. Follow - @starmexxx I'll keep finding them before they close //

Prompts

┌──────────────────────────┬───────────────┬──────────────┐
│ Subscription             │ Monthly cost  │ Annual cost  │
├──────────────────────────┼───────────────┼──────────────┤
│ Claude Code Max (20x)    │ $200/month    │ $2,400/year  │
│ ChatGPT Pro              │ $200/month    │ $2,400/year  │
│ Gemini Advanced          │ $20/month     │ $240/year    │
│ Cursor Pro               │ $20/month     │ $240/year    │
├──────────────────────────┼───────────────┼──────────────┤
│ Total heavy user         │ $440/month    │ $5,280/year  │
└──────────────────────────┴───────────────┴──────────────┘

┌────────────────────────┬──────────────────────────┐
│ Spec                   │ GMKtec EVO-X2             │
├────────────────────────┼──────────────────────────┤
│ Chip                   │ AMD Ryzen AI Max+ 395     │
│ Cores / Threads        │ 16 / 32                   │
│ Max clock              │ 5.1 GHz                   │
│ GPU                    │ 40 RDNA 3.5 CUs           │
│ NPU                    │ 50 TOPS                   │
│ Combined AI perf       │ 126 TOPS                  │
│ Unified memory         │ up to 128GB               │
│ Usable VRAM (Linux)    │ up to 110GB               │
│ Price                  │ $1,700 - $2,000           │
└────────────────────────┴──────────────────────────┘

┌──────────────────┬────────────┬───────────────────────────┐
│ Model            │ VRAM needed│ Result on EVO-X2           │
├──────────────────┼────────────┼───────────────────────────┤
│ Qwen3-235B       │ ~110GB     │ Runs fully, smoothly       │
│ DeepSeek-V3      │ ~100GB     │ Runs comfortably           │
│ Llama 3.3 70B    │ ~42GB      │ Fast, plenty of headroom   │
│ Qwen3.6 27B      │ ~16GB      │ Very fast, daily driver    │
└──────────────────┴────────────┴───────────────────────────┘