AI Model Compression: From Frontier to Mobile in 23 Months

Advanced AI models that required 1.8 trillion parameters two years ago now run on smartphones with only 4 billion parameters, a 450x compression achieved through better algorithms, talent concentration, and massive capital investment. Google's Gemma 4 E4B now matches GPT-4o performance on mobile devices, with further releases expected from DeepSeek, Qwen, Kimi, and Minimax in coming weeks. At current compression rates, frontier-level AI capabilities will run on consumer phones before they require upgrades.