Three desktops on AMD’s Ryzen AI Max+ 395 (“Strix Halo”): the GMKtec
EVO-X2 ($1,999, 2025), the Framework Desktop ($1,999, 2025), and the AMD
Ryzen AI Halo reference box ($3,999, 2026). All three share the chip’s
specs - 128GB unified LPDDR5X at 256 GB/s, 120W, Radeon 8060S graphics -
and differ in enclosure, cooling, and warranty.
What it does well:
-
128GB at $1,999: half the DGX Spark’s price for the same capacity,
if you accept 256 vs 273 GB/s.
-
x86 compatibility: full ROCm and Vulkan runtimes, no ARM
translation layer for exotic wheels or CUDA-alternative tooling.
Where it falls short: 256 GB/s is the ceiling, so single-stream
decode trails NVIDIA discrete cards by a wide margin; ROCm support for
consumer-class APUs has trailed CUDA in every runtime.
Run it locally: Ollama and LM Studio (Vulkan), llama.cpp with ROCm.
The modeldex results rank checkpoints by
fit and estimated tok/s - memory fit is identical across all three
enclosures.