Google Gemma 3 — Open Weights, 128K Context, and the QAT Advantage

Google's Gemma 3 family (March 2025) delivers a 27B model that outscores LLaMA 3.1 405B on LMArena, four sizes from 270M to 27B, vision-language capabilities on three of those sizes, and official quantization-aware training (QAT) variants for consumer GPU deployment. Gemma Terms license, not Apache 2.0. We review the full family.