Deem-Yangs: 95% Leaner MoE Inference via KV Bank

Yang, D.'s paper solves the problem of compiling knowledge from a large model (e.g., DeepSeek-V4) and injecting it as full-layer KV into a local small model (e.g., Qwen3-0.6B, Qwen3.8-27B). The present paper takes this working mechanism and applies it to a single merged model. It proposes the Deem-Yangs (Model) and the MoE-YS Mechanism, which compiles the large model's (e.g., DeepSeek-V4-Pro) own routed-expert computation into per-layer static KV banks; a small in-model subnetwork (the shared expert + active routed core) remains resident, while 98% of the parameters exist only in KV form, thereby saving 95% of memory usage. The architecture is grounded in four reproduction-verified findings (F1-F4) of the author's preceding full-layer memory injection work (DOI 10.5281/zenodo.22305403).