LLaMA 3 70B Llamafiles

What's the cheapest hardware setup that can run a 70B model at tolerably interactive rates? (say 10 characters a second)

What’s the relationship between llamafile and llama.cpp these days? IIRC, it was originally llama.cpp compiled with Cosmopolitan Libc. I’ve read more recently of optimizations in llamafile with comparisons against llama.cpp. Are those pushed upstream? Have the projects diverged?

8B llamafile when? :^)