It is distributed with an Apache 2.0 license that allows commercial use

Google has introduced Gemma 4, a completely open new frontier model designed to run locally on Android devices and computers, with a size reaching 31 billion parameters.

Gemma 4 is the new iteration of Google’s open model family, which has been designed for advanced reasoning and agent-based workflows. To do this, its capabilities include deep logic and multi-step planning.

Google has highlighted that it offers “an unprecedented level of intelligence per parameter”, which means that it “achieves cutting-edge capabilities with significantly less hardware overhead”, as noted on its official blog.

Gemma 4 comes in four sizes: Cash 2 billion parameters, Cash 4 billion, Expert Mix 26 billion and Denso 31 billion; with a context window of up to 256,000 tokens.

The new models also natively process audio and video, and can generate code without an internet connection thanks to their focus on local processing. They also have support for more than 140 languages.

The company has highlighted its open nature, which it shares with the rest of the Gemma family, but has insisted that it is distributed with an Apache 2.0 license, which “provides a basis for total developer flexibility and digital sovereignty, giving you absolute control over your data, infrastructure and models,” Google said. In this way, the model can be modified locally for personal or commercial purposes.

Gemma 4 can run locally on “billions of Android devices worldwide” and “GPUs from laptops to workstations to developer accelerators.” It is now available on Google AI Studio and on Hugging Face, Kaggle and Ollama.