It's a massive performance upgrade, which would make current sized models better and tiny phone-sized models viable. Only problem is that models need to be retrained to use it and afaik, no one significant has done it yet.
Nice. Any additional info on how difficult it was to train this and whether we can expect more? They have a 3B model in the demo video, but doesn't seem like they released that... I mean I'd like something a bit larger.
6 Comments
WildPalmTree@lemmy.world · 7 pts · 1y
ELI5 1-bit module. With three attempts, i got nothing out of it, so I assume it's a simpler, more energy efficient model.
icecreamtaco@lemmy.world · 5 pts · 1y
It's a massive performance upgrade, which would make current sized models better and tiny phone-sized models viable. Only problem is that models need to be retrained to use it and afaik, no one significant has done it yet.
thickertoofan@lemm.ee · 4 pts · 1y
i'm not the smartest out there to explain it but it's like ...instead of floating point numbers as the weights, its just -1,0,1.
simple@lemm.ee · 4 pts · 1y
This wasn't out? I've been hearing about BitNet for a while, just that there wasn't a good 1-bit model out there.
thickertoofan@lemm.ee · 3 pts · 1y
it was, it's just that they have officially released a 2B model trained for the BitNet architecture
hendrik@palaver.p3x.de · 1 pts · 1y
Nice. Any additional info on how difficult it was to train this and whether we can expect more? They have a 3B model in the demo video, but doesn't seem like they released that... I mean I'd like something a bit larger.