> This sounds like what they call "Bamba-9B" is actually an 18B model quantised to 8 bits. No it doesn't? The fact that it is 18 GB with 16 bit per parameter b

> This sounds like what they call "Bamba-9B" is actually an 18B model quantised to 8 bits. No it doesn't? The fact that it is 18 GB with 16 bit per parameter before quantization means that it is a 9B parameter model.