Gemma 4

Apache 2.0 licensing is the headline of Gemma 4, the first release in Google DeepMind's open-weight line to drop the custom Gemma terms that complicated commercial use. The April 2026 family spans four sizes: efficient E2B and E4B variants built for phones and edge hardware, a 26B mixture-of-experts model, and a 31B dense flagship that debuted as the third-ranked open model on the LMArena text leaderboard with the 26B in sixth. All sizes accept images alongside text, the smaller variants take native audio input, and context windows reach 128K on the small models and 256K on the larger ones, with coverage of more than 140 languages. Weights download from Hugging Face, Kaggle, and Ollama and run through transformers, llama.cpp, and the usual local stacks; the E2B and E4B models fit consumer hardware without a discrete GPU, while Google serves the whole family through AI Studio and Vertex AI for hosted use. For teams that skipped earlier Gemma releases over licensing friction, this generation removes the objection entirely.