if you start excluding the 1000+ B param models …
using smaller models, would initially ease hardware demand by 60% .
OFC you cant… and probably shouldnt, ignore and disrespect SOTA flagship models
Even “big” open source models like DSV4 and Ling/Ring are very efficient. They’re big, but (seemingly) sparser than US models, so they’re cheap.
They run surprisingly well with hybrid CPU+GPU inference on desktops. And thats not even getting into the efficient attention mechanisms.
I can run DSV4 Flash, barely quantized, with ~1M context on my Ryzen desktop at ~11 tokens/s. If you told me that two years ago, I would not have believed you.
if you start excluding the 1000+ B param models …
using smaller models, would initially ease hardware demand by 60% .
OFC you cant… and probably shouldnt, ignore and disrespect SOTA flagship models
Even “big” open source models like DSV4 and Ling/Ring are very efficient. They’re big, but (seemingly) sparser than US models, so they’re cheap.
They run surprisingly well with hybrid CPU+GPU inference on desktops. And thats not even getting into the efficient attention mechanisms.
I can run DSV4 Flash, barely quantized, with ~1M context on my Ryzen desktop at ~11 tokens/s. If you told me that two years ago, I would not have believed you.