Google has launched Gemma 4, a new family of open-weight AI models that the company describes as its most capable to date. The models are built on the same research and technology as Google DeepMind’s proprietary Gemini 3 system and are released under an Apache 2.0 open-source license, which allows developers to use and modify them freely for commercial purposes.
The Gemma 4 family comes in four sizes: two compact models called E2B and E4B designed for mobile devices and edge hardware, and two larger models of 26 billion and 31 billion parameters designed for personal computers and developer workstations. According to Google, the 31B model currently ranks third among all open AI models on the Arena AI text leaderboard, while the 26B model ranks sixth. Google claims both models outperform competing models that are up to 20 times larger.
The E2B and E4B models are built specifically for low-power environments. They can run fully offline on smartphones, Raspberry Pi devices, and similar hardware. Both support audio and video input in addition to text. The larger models support context windows of up to 256,000 tokens, meaning they can process the equivalent of several long documents in a single session.
Key capabilities across the model family include:
- Advanced reasoning for mathematics and multi-step logic
- Agentic workflows with support for function calling and structured output
- Code generation for offline development environments
- Multimodal input including images, video, and audio
- Support for over 140 languages
The Apache 2.0 license is a notable change from licensing terms used in earlier Gemma generations. Clément Delangue, co-founder and CEO of Hugging Face, called the licensing decision “a huge milestone.” The models are available through Hugging Face, Kaggle, Ollama, and LM Studio on the day of launch.
Google states that the models have been tested under the same security protocols applied to its proprietary systems. The company says this is intended to make Gemma 4 suitable for enterprise and government use cases where data control and regulatory compliance are important.
The release also includes tools for customization. Developers can fine-tune the models using platforms such as Google Colab, Vertex AI, and standard frameworks including Keras and JAX. Google highlights two existing examples of this: a Bulgarian-language model called BgGPT developed by INSAIT, and a cancer research project called Cell2Sentence-Scale developed in collaboration with Yale University.
Since the first Gemma models launched, Google says developers have downloaded the series more than 400 million times and created over 100,000 variants.
Sources: Google Blog, Google DeepMind
