Google Releases Gemma 4 QAT Variants for On-Device Inference

Google / Price Per Token·medium signal

On June 5, Google shipped quantization-aware-training (QAT) versions of Gemma 4 aimed at running efficiently on mobile and laptop hardware, optimizing model compression for local deployment. This follows the Gemma 4 12B launch (77.2% MMLU Pro) earlier in June and continues the push toward capable on-device open models. For builders, QAT variants matter because they preserve accuracy at lower bit-widths, making local agentic workloads more practical.