Gemma 4 models now use quantization-aware training to use less memory while retaining quality performance.