run huge MoE models on own hardware

A deeply technical MoE inference engine a solo dev cannot realistically replicate, facing a dominant incumbent (llama.cpp) — a poor standalone product bet despite strong stars.