I've got a 48c Epyc with 2x3090s and 512gb ddr4 3200. It's good enough for 25+ tps with deepseek so I'm hoping for similar performance with less overthinking.

This is needs ~80GB of fast memory at 4 bits per weight. Faster memory is better, but probably even something like 3090 + 64GB RAM should work (not fast, but maybe even 20-30 t/s? llama.cpp support pending).