Google has released Gemma 4, a new family of open-weight AI models built on the technology behind its proprietary Gemini 3 Pro. The company is offering four variants with different parameter counts: 2-billion and 4-billion models for edge devices, plus 26-billion and 31-billion models for more powerful hardware. Google claims these systems deliver "unprecedented intelligence-per-parameter" performance.

To prove its point, Google highlights that the 31-billion and 26-billion Gemma 4 models rank third and sixth on Arena AI's text leaderboard, outperforming models 20 times their size. All four models can process video and images for tasks like optical character recognition, while the smaller variants add audio and speech capabilities. The models also support offline code generation in over 140 languages.

Google is releasing Gemma 4 under an Apache 2.0 license, a shift from the proprietary Gemma license used for previous versions. This change gives developers complete freedom to modify and deploy the models across any environment. The model weights are available through Hugging Face, Kaggle, and Ollama for anyone wanting to test the systems themselves.