Google DeepMind released Gemma 4 quantization-aware training checkpoints, which optimize compression for better mobile and laptop efficiency.

These checkpoints use Quantization-Aware Training (QAT) to simulate quantization during training, minimizing quality loss when compressing the model. The release includes QAT checkpoints for the Q4_0 format and a new format for mobile, reducing the memory footprint of Gemma 4 E2B to 1GB. QAT integrates quantization into training, enhancing quality beyond standard Post-Training Quantization (PTQ) while maintaining Gemma 4's capabilities.

Source: blog.google