Hostinger finds MoE merging breaks Qwen3-30B tasks
Researchers at Hostinger testing AI model compression found that while pruned and merged Mixture-of-Experts models scored similarly on standard benchmarks, merging caused significant failures in real-world product tasks such as breaking email templates on mobile screens, generating endless text loops, and mixing languages mid-sentence. Pruning outperformed merging across most test matchups, leading the team to recommend pruning over merging, limiting expert cuts to twenty-five percent for text-writing models, and validating compressed models on specific product constraints rather than relying on benchmarks. Open article →