LLAMA 3 and Phi-3 came out

[sebastienbo]

Hi,

Two important LLM's have come out to replace their version 2

LLAMA 3 and Phi-3 , when will they be available on llama file?

The main readme still points to llama 2 and phi-2

[sebastienbo]

Phi-3 just came out on llamafile, but why is it so big?
On hugging face Phi-3 mini is only 2GB .
On windows an executable cannot be more then 4GB, it would have been nice to have a single executable for phi-3

Who can explain why it is 7gb here ?

[normatovjj]

That's because it's unquantised, in BF16 format meaning every weight/parameter is saved in 2 bytes. The model itself has 3.82B parameters, times 2 bytes, you get ~7.65GB model