Solid analysis, these quotes were particularly damning imo:
"You don’t accidentally forget to disclose that your memory is split into two non-unified pools. That’s a deliberate omission of the single most important architectural limitation of the device."
"Read that filename. layer_27_36. That’s not dynamic hot neuron routing, that’s static layer assignment"
To the author’s point, I was able to run Qwen3.5 122B at close to 20 tokens/s on a Ryzen AI Max 395+ mini PC with 128GB (64/64 split, for some reason it needs a lot of system RAM).
Based on the exploded views I found and some text references in a product page or something I can't quite remember it uses a vapor chamber and a dual fan configuration. If you go frame by frame on the exploded view animation it can be seen.
"You don’t accidentally forget to disclose that your memory is split into two non-unified pools. That’s a deliberate omission of the single most important architectural limitation of the device."
"Read that filename. layer_27_36. That’s not dynamic hot neuron routing, that’s static layer assignment"
And that's at idle!! (15w) https://www.jeffgeerling.com/blog/2025/minisforum-stuffs-ent...
Now add two NPUs. Personally I'm for it. But it is going to need quite the cooling. A bunch of Frore Airjet mems devices?