Gemma 4 QAT models: Optimizing compression for mobile and laptop efficiency

Google DeepMind has released Quantization-Aware Training (QAT) optimized versions of Gemma 4 models designed to reduce memory requirements and enable efficient on-device performance on mobile devices and consumer GPUs. The new checkpoints include support for the popular Q4_0 quantization format and a novel mobile-specialized quantization schema that reduces the Gemma 4 E2B model's memory footprint to under 1GB while preserving model quality. The mobile optimization uses techniques like static activations, channel-wise quantization, and targeted 2-bit compression to maximize efficiency on edge hardware.