Interesting, I really would love to see Llama 405B again. It has a lot of knowledge that doesn’t necessarily reflect on benchmarks. Due to it’s size of course, but even more so because it wasn’t just trained on the select topics that would only score well. Maybe you know what I mean already, it just knows a lot of random details and information that other models don’t.
I looked at the “On Request” page but it isn’t clear how it would work, a dedicated node sounds expensive, is that by the hour? Or is there another way to just access that model again?