[wruza]

Oh, I have a question, maybe you know.

Assuming the same model sizes in gigabytes, which one to choose: a higher-B lower-bit or a lower-B higher-bit? Is there a silver bullet? Like “yeah always take 4-bit 13B over 8-bit 7B”.

Or are same-sized models basically equal in this regard?

[anon373839]

I would say 9 times out of 10, you will get better results from a Q4 model that’s a size class larger than a smaller model at Q8. But it’s best not to go below Q4.

[nenaoki]

My understanding is that models are currently undertrained and not very "dense", so Q4 doesn't hurt very much now but it may in future denser models.

[anon373839]

That may well be true. I know that earlier models like Llama 1 65B could tolerate more aggressive quantization, which supports that idea.