The industry keeps benchmarking against other labs instead of against user needs: If a 3.5GB model answers 95% of everyday questions well enough, the remaining few percent has to justify hundreds of gigabytes of weights, huge energy bills and constant cloud costs.