Gemma 4 QAT models shrink to 1 GB for mobile‑grade AI
Google released Gemma 4 models trained with quantization-aware training (QAT), cutting the 2B model’s memory to 1 GB and improving edge-device performance. The new mobile‑specialized quantization schema uses static activations, channel‑wise and 2‑bit compression to preserve quality while enabling fast inference on phones and laptops.