The architecture of open-weight LLMs: MoE, hybrid layers, ternary weights