Strix Halo mini PCs put 128GB of unified memory within reach for local AI
Originally announced by AMD . We link the primary source so you can read it for yourself.
A wave of compact mini PCs built on AMD’s Ryzen AI Max+ 395, codenamed Strix Halo, has made a new kind of local AI machine practical: one with up to 128GB of unified memory, of which a large portion can be assigned as VRAM, in a small, low-power box.
What it is
The Ryzen AI Max+ 395 pairs a capable integrated GPU with unified memory shared between processor and graphics. On the 128GB configuration, much of that memory can serve models directly. That lets these machines hold models that will not fit on any single consumer discrete card, such as a 70B at higher precision, at a fraction of the power a multi-GPU rig would draw.
The trade-off
The catch is bandwidth. Memory bandwidth on these machines is well below a high-end discrete GPU, so while they can hold large models, they generate more slowly than a card of similar memory would. This is the same trade Apple Silicon makes: capacity and efficiency in exchange for speed. For interactive use, a mixture-of-experts model like Qwen3 30B-A3B, which activates only a fraction of its parameters per token, is a natural fit.
Why it matters
This is a genuinely new option in the local AI hardware landscape, sitting between a consumer GPU and a high-memory Mac. For people who want to run large models at home quietly and cheaply, and who can accept slower generation, it widens the field beyond NVIDIA and Apple. As always, our cost calculator and hardware matrix can help you weigh it against the alternatives.