MoE model benchmark gains 32% throughput by reordering expert weights

A developer's benchmark showed reordering expert weights in a Mixture-of-Experts model by co-activation reduced disk reads 2.23× on an 80B model. Independent testers then measured +32.3% decode throughput and -26.3% time-to-first-token on a 235B model running on a 48GB MacBook. The process involved sharing data with inference engine maintainers, who refuted two of the original three pitches with their own measurements.