Comment by zkmon

I don't get it. It's file size is about 6 times larger than 27B model for the same quant, but the performance improvement is hardly 10% across all benchmarks, according the metrics on it's hf page. Why should one devote so much more hardware for so little benefit?