Smaller AI models don't save phone battery, but a different architecture does

The paper measured how much energy large AI models use when running on a phone. It turns out that making models smaller to fit on a phone does not save much battery life; instead, the way the model is built, especially using a "Mixture-of-Experts" design, is what actually makes it energy efficient.

Everyone assumed that if you could just shrink a big AI model, it would run efficiently on a phone. This paper shows that assumption is wrong for battery life. Mobile phone makers and chip designers now have clear data that points to specific model architectures, like Mixture-of-Experts, as the real path to powerful AI that doesn't drain your battery. This changes how they will design future chips and software for on-device AI.