About a year ago, people were publishing a lot of benchmarks about various quants of models. I understand that it is not really feasible with the current (and other welcome) frequent releases of new models, but on the other side, it may be still useful to know locally whether q3 of this model is better than q6 of that model. I've checked a few benchmarks, but it seems they are versatile, and the models may generate millions of tokens, which, with a 300b+ moe model on a home setup of 10-20 t/s ...