Google DeepMind released Gemma 4 on April 2, a new generation of open-weight AI models built on the research behind its Gemini 3 series. The models are published under the Apache 2.0 license, which lets developers use, modify and ship them commercially without the custom usage terms that governed earlier Gemma releases.
The launch comes as open models from Chinese developers such as DeepSeek, Alibaba’s Qwen team and Moonshot have gained wide adoption, and it is widely read as Google’s effort to keep Western open models competitive in that race.
Four sizes, from servers to phones
Gemma 4 ships in four sizes. According to Google, the largest variants are designed to run on a single Nvidia H100 GPU, making them accessible to organizations that cannot afford multi-GPU clusters, while the smallest are built for phones, Internet of Things devices and other low-power hardware. Google says the top model delivers performance that is strong for its parameter count, the “byte for byte” claim in the company’s announcement, though independent benchmarks will determine how it compares with rivals.
All of the models accept multiple input types. Google says Gemma 4 can process text, images, audio and video, and that the models were trained for agentic use, meaning they can follow multi-step instructions, call external tools and functions, and generate code. Training covered more than 140 languages, and the context window, the amount of text the model can consider at once, is larger than in Gemma 3.
A large existing community
Google says earlier Gemma models have been downloaded more than 400 million times and that the community has published over 100,000 variants fine-tuned for specific purposes, including medical, educational and regional-language applications. That installed base is one of the main reasons developers pay attention to a new Gemma release: tooling, tutorials and deployment paths already exist.
The models are available through Hugging Face, Kaggle, Google’s Vertex AI platform and popular local inference tools, and the smaller variants can be tried directly on a phone through Google’s AI Edge Gallery app.
Why the license matters
The move to Apache 2.0 is arguably as significant as the capability gains. Previous Gemma models used a Google-specific license with usage restrictions that made some companies cautious about building on them. A standard permissive license removes that friction and puts Gemma on the same legal footing as many of the Chinese open models it competes with. For developers, the practical result is a capable, multimodal model family they can run anywhere from a data center to a handset without licensing uncertainty.
