Developer's AI model optimization validated by strangers, yielding 32% throughput gain

A developer working on a project called mbolt discovered that reordering expert weights inside a Mixture-of-Experts (MoE) model file based on co-activation patterns could significantly reduce disk reads per token. After initial solo testing showed a 2.23× read reduction, the developer took the work directly to maintainers of relevant inference engines rather than posting publicly. An independent researcher tested the approach on a 235-billion-parameter model running on a 48GB MacBook and recorded a 32.3% increase in decode throughput and a 26.3% reduction in time-to-first-token. During the collaborative review process, two of the developer's three original optimization pitches were disproved using real engine data, which the developer describes as the most productive part of the process. The article argues that targeting a small number of highly relevant engineers and inviting rigorous external testing is a more effective validation strategy than broad public announcements.