>>105851704
There are some ML researchers that do quantization work but they are few compared to other fields in ML right now.
>>105851861
Doesn't matter, they can say anything from 128k to 1m and it wouldn't actually be true since even their closed sourced models have scores in NoLiMa that are around 16k. Unless we see this score the same, it's effectively useless, not to mentioned probably slopped to hell.