Google releases Gemma 4 open weight AI models with Apache 2.0 license

Google has released Gemma 4, a family of open weight AI models based on technology from Gemini 3. The lineup includes four variants: 2 billion and 4 billion parameter models for edge devices such as smartphones, along with 26 billion Mixture of Experts and 31 billion Dense models for more powerful hardware.

Gemma 4 is also designed for multimodal tasks. All models support image and video input for use cases like optical character recognition, while the smaller versions also handle audio input and speech understanding. Google says the models were trained in more than 140 languages, giving the family broad multilingual support.

Google also claims Gemma 4 offers strong performance relative to its size, noting that its two largest models recently placed third and sixth on Arena AI’s text leaderboard while outperforming much larger systems. The models can also generate code offline, and Google is now releasing them under an Apache 2.0 license instead of its previous custom license, giving developers more freedom to modify and deploy them.