Models at this size are regularly quantized to 5/4bit to reduce the size. While there is some degradation, it isn’t as substantial as you would expect.