[int_19h]

When people tried 3-bit quantization for 7B models before, it did not exactly go well in terms of detrimental side effects. Are you using some new quantization techniques that mitigate that?