Are we in the era of ternary and binary models yet?

Given the cost of training and the speed at which advancements are coming out, even if bitnet were a considerable improvement it wouldn't be economically sensible to train a new model from the ground up on that architecture. Just because SotA doesn't use it, doesn't mean it's not better. It's however not better *enough* to justify investing into it compared to increasing investment on what is already available. Imo we'll see new models that combine a bunch of this kind of improvements between this year and the next.