> https://github.com/ggerganov/llama.cpp/issues/1602#issuecomm...

Here somebody quantized it down to 29929.56MB .