mobe-mixture-basis-experts

Compresses MoE language models through shared basis factorization of expert weight matrices, achieving 24-30% parameter reduction with minimal accuracy loss.

Independent GitHub snapshot · Automated classification · Original documentation governs