>>79510096Probably won't ever happen, requires hundreds of dollars in capital to rent the training to do that. I should've been more clear when talking about SOTA, I meant at smaller sizes (< 10B parameters). Gap is closing though on the smaller models, it's close enough to try but I am not going to try Gemma again until things get fixed in the inference software.
>>79558725Yes, but the writing can be stiffed and you want some other variety. OpenRouter will probably have Gemma 2 cheap once things are working.
