The video explains that the 2x speed improvement in running local AI models like Quen 3.8 on Apple Silicon Macs using Splash is mainly due to speculative decoding, which leverages a smaller draft model to predict multiple tokens at once, reducing memory bandwidth bottlenecks. While Splash’s optimized inference engine contributes, the key to faster performance lies in multi-token prediction techniques that enhance real-world responsiveness, especially on high-memory Macs, though accessibility is limited by proprietary formats and hardware requirements.