Overview
- Google DeepMind, which unveiled Gemma 4 on Thursday, released open weights under the commercially permissive Apache 2.0 license.
- The lineup spans E2B, E4B, a 26B Mixture‑of‑Experts model, and a 31B dense model, with Google citing Arena AI ranks of #6 and #3 for the 26B and 31B.
- Built with the Pixel team plus Qualcomm and MediaTek, the smaller E2B and E4B variants run fully offline with near‑zero latency on phones, Raspberry Pi, and Jetson Orin Nano.
- All models process images and video, the edge pair adds audio input, context windows range from 128K to 256K tokens, and training covers more than 140 languages.
- Developers can try the models in Google AI Studio and the AICore Developer Preview or download weights from Hugging Face, Kaggle, and Ollama, with day‑one support in Android Studio plus NVIDIA and Docker integrations.