Hacker News new | ask | show | jobs
by vletal 552 days ago
That would be 1.38 bits per weight on average, which I can confidently guess would not perform well.
2 comments
qeternity 552 days ago
BitNet is functional at 1.58 bpw.
link
Lerc 552 days ago
The model card says the 70B is 16 bit so I think you have twice that
link