> If the model was designed specifically to quantize down to 1.58b, then it's different.

> AFAIK, there's no large models designed for this yet.

Isn't BitNet b1.58 2B4T what you are looking for? (haven't tried it myself though)