Google's Gemma 4, ranging from a 2-billion-parameter model to a 31-billion-parameter variant, targets offline, on-device assistants and reasoning tasks. Google describes it as ‘byte for byte, the most capable open models’, maximizing ‘intelligence per parameter’. Harshil Shah, Head of Generative AI at R Systems, notes that Gemma shifts the paradigm with efficiency, particularly the 12B model, which uses a unified, encoder-free multimodal architecture that feeds raw image and audio representations directly into its core language network, reducing memory overhead. Shah identifies on-device, multimodal workflows as the definitive sweet spot. Gemma 4’s model card lists document/PDF parsing, screen and UI understanding, OCR, and variable-resolution image processing as core capabilities. A configurable thinking mode, enabled by a dedicated control token, improves reasoning on mathematics, coding, and multi-step problem-solving. Quantization-Aware Training (QAT) enables aggressive compression without significantly degrading capabilities, while Multi-Token Prediction (MTP) speeds up inference by using a lightweight draft model to predict multiple tokens ahead before the main model verifies them.