560k+ downloads last month
Yo what the heck, 560k downloads in a month is nuts lmao.
haha yes, I was very surprised by that number myself!
the weird thing is that its not even the most capable pepe (that would obviously be the 70B version).
and the downloads are the fp16 weights, so this highly suggests its corpo downloads that automatically pulling it via some pipeline.
i wonder what they do with it, but we'll likely never know :')
If the corpo uses Pepe to write e-mails they are going to be glorious. Bro will roast everyone.
It's weird to still download L3 models when all the FOMO is on G4 and Qwen lately, and uncensored version of them exist on HF. There has to be some unique use for Pepe 8b we don't know.
those would be a very entertaining emails to read for sure ^^
regarding gemma4 and the new qwen-- they are x100 better BASE MODELS for sure, but i feel that the latest generation of models are so densely packed with intelligence-- that it kinda makes them brittle.
most attempts to tune them often destroys other parts (its always the case in any finetune, but its getting worse every gen, with the exception of llama-1 and llama-2 era, when models were VERY undertrained). pepe 8b outperforms the base model which is rare for such an extreme finetune (extreme in terms of how radically different the output vs the base model).
and you are for sure 100% right about the last point, someone is definitely doing something with pepe at scale, and i truly wish i could know exactly what... the curiosity is real haha
idk i still find assistant pepe more useful than gemma 4 overall. G4's intelligence is quite narrow, maybe better for STEM but not RP. my finetune attempts of it so far have resulted in similar findings as yours.
if @xai-org ever releases grok 3 weights you should feed them pepe dataset