Thilo Michael

Congratulations, these results look really promising. However, I really don't think the comparison with FP-16 versions of other small models is fair. At least use the 8-bit or even 4-bit quantized version to give a more accurate performance and to reflect the versions people actually use.