The model is too large for 32GB
You want a quiet, self-contained box for private inference, and capacity matters more than interactive generation speed.
Local AI · capacity machine
GMKtec EVO-X2 for local LLMs
A Ryzen AI Max+ 395 mini PC with one unusual advantage: up to 128GB of unified memory. Buy it for models that do not fit on a 32GB consumer GPU—not because an integrated GPU is secretly faster than one.
Best for Private 70B–120B-class modelsNot for Maximum tokens per secondPrice checked 2026-07-09
July 2026 buying guide · prices update daily · specifications linked to first-party sources
Your model already fits an RTX card
For 7B–32B models, a high-end discrete GPU's much wider memory bus is usually the more performance-oriented choice.
What fits: 64GB versus 128GB
AMD documents up to 48GB dedicated graphics memory on a 64GB system and up to 96GB on a 128GB system. Our table applies conservative weight-plus-headroom estimates; real limits vary with context length and runtime.
| Model | Total params | INT4 need | 64GB model (48GB VGM) | 128GB model (96GB VGM) |
|---|---|---|---|---|
| GPT-OSS 120BApache-2.0 | 117B | ~81 GB | no | fits |
| Llama 4 ScoutLlama 4 Community | 109B | ~75 GB | no | fits |
| Llama 3.3 70BLlama 3.3 Community | 70B | ~49 GB | no | fits |
| Qwen3 32BApache-2.0 | 32.8B | ~23 GB | fits | fits |
| Gemma 3 27BGemma | 27B | ~19 GB | fits | fits |
| Phi-4MIT | 14.7B | ~11 GB | fits | fits |
“Fits” means the estimated weights plus ~25% KV/activation headroom stay inside the dedicated graphics allocation. It is not a speed benchmark or a guarantee for every context length.
Editable economics
Buy this box or rent a cloud GPU?
Compare cash cost only. The calculator automatically excludes cloud cards that cannot hold the selected memory target. It does not pretend the hardware has equal throughput.
Price record
The first-party product feed is checked with the site's daily compute-price pipeline. We keep one observation per day so temporary discounts become visible instead of disappearing.
| Configuration | Today | List / compare-at | Availability | Days recorded |
|---|---|---|---|---|
| 64GB RAM + 1TB SSD | $2000 | $2200 | In stock | 1 |
| 128GB RAM + 2TB SSD | $3400 | $4000 | In stock | 1 |
The hardware facts that matter
ProcessorRyzen AI Max+ 39516 Zen 5 cores · 45–120W cTDP
Integrated GPURadeon 8060S40 RDNA 3.5 compute units
Memory64GB or 128GB LPDDR5x-8000256-bit · 256 GB/s
AI allocationUp to 48GB or 96GB VGMDepends on installed system memory
Verify the processor and memory details in AMD'sofficial specificationand its Variable Graphics Memory guide. This is an evidence-based buying guide, not a hands-on review; vendor benchmark claims are not used as editorial performance conclusions.
FAQ
Can it run a 70B model?
The 128GB version has the room for many 70B models at 4-bit quantization. Our conservative estimate is ~49GB before longer-context growth, just over the 64GB model's 48GB VGM allocation.
Why not just buy an RTX 5090?
If your model fits in 32GB, that is the sharper performance question. The EVO-X2 earns its place when the model does not fit; its advantage is memory capacity, not raw bandwidth.
Does 128GB mean 128GB of VRAM?
No. It is unified system memory. AMD documents up to 96GB assigned as dedicated graphics memory, with some additional shared access possible. The CPU also needs memory.
Is the current price guaranteed?
No. The displayed price and stock state are a daily snapshot from GMKtec's product feed. Confirm the final price, plug type, shipping and taxes on the retailer page.
The GMKtec link is an affiliate link; llmhosting.ai may earn a commission if you buy. It does not affect the model-fit math or the comparison. How we make money ·all local hardware · API vs rented GPU calculator