Google Launches Gemma 4 Open-Weight Models Under Apache 2.0, Runs Frontier AI on a Single H100 GPU

Google DeepMind released Gemma 4, a family of four open-weight multimodal models built on the same research as Gemini 3, now available under the permissive Apache 2.0 license — a first for the Gemma line and a significant liberalization from previous terms that allowed Google to revoke access. The 31B Dense variant ranks third on the Arena AI Text leaderboard among open models, all variants fit on a single 80GB NVIDIA H100 GPU, and smaller variants reportedly outperform models 20x their size. The models support over 140 languages and image, text, and audio inputs, are deployable on NVIDIA RTX GPUs, Google AI Studio, and AI Edge Gallery, and the llama.cpp community rapidly merged support. Hugging Face co-founder Clément Delangue called the Apache 2.0 switch 'a huge milestone' for enterprise sovereignty, and NVIDIA announced optimized Gemma 4 deployment support on its RTX consumer GPU line.