Google has released the Gemma 4 open AI models, replacing its predecessor, Gemma 3, which launched over a year ago. The new models come in four sizes designed for local usage, with the large 26B Mixture of Experts and 31B Dense designed to run unquantized in bfloat16 format on a single 80GB NVIDIA H100 GPU. Google has also removed the custom Gemma license and will use Apache 2.0 instead of the current version of the open-weight model. These new models reportedly use less memory and battery than their predecessors, and Google claims they will be more capable on local hardware.