[nok22kon]

paper is a bit old, but matches current empirical recommandation: a good starting point is the biggest model you can fit at 4 bit

https://arxiv.org/abs/2212.09720