Gemma 3 dazzles with its knack for multimodality, accepting both text and vision-language inputs and accommodating context windows up to 128k tokens. With multilingual prowess spanning 140 languages, it showcases exceptional math, reasoning, and chat capabilities. Available in four flavors: 1B, 4B, 12B, and 27B, it's a real feast for tech enthusiasts. Trained on Google TPUs, it masters distillation and reinforcement learning, clinching a top score of 1338 in LMArena. An adaptive window allows it to handle high-resolution images, and a SigLIP-based vision encoder guarantees uniform image analysis in all sizes.