On-Device Models: Not Cloud Downsizing, but Physical World Syntax Reconstruction

Last year, when I used AlphaFold2 to predict the structure of a mutated protein, a single inference took 18 minutes, and the GPU fans were spinning so hard they seemed ready to take off. This year, I tried using the same model on a mobile device—it wasn't just too slow; it simply wouldn't run at all. According to a review I saw on bioRxiv last year, 90% of protein structure prediction tasks cannot run in real-time on mobile devices.