4 ms·

The 4bit quantized 13B models, give really decent answers (not as good as gpt4, but often as good as gpt 3)