Google released Gemma 4 on April 2, its first major update to the open model family in a year. The release includes four model variants and, critically, a license change: Gemma 4 is distributed under the Apache 2.0 license, one of the most permissive open-source licenses in software. [1]

The previous Gemma models were released under a custom Google license that restricted commercial use. Apache 2.0 removes those restrictions. Anyone can use, modify, distribute and sell Gemma 4 without asking Google for permission. [2]

The Models

Gemma 4 includes a 31-billion-parameter dense model and several smaller variants optimized for edge deployment. The models support 256,000-token context windows — a significant increase over prior versions and competitive with leading proprietary systems. [3]

The models are multimodal, capable of processing text, images and structured data. They are designed to run on hardware ranging from cloud GPUs to mobile devices, making them accessible to developers who cannot afford the infrastructure required for larger models. [4]

The License

The shift to Apache 2.0 is the most consequential aspect of the release. Google's previous Gemma license required users to comply with Google's acceptable use policy and restricted certain commercial applications. Apache 2.0 imposes no such conditions. It is the same license used by TensorFlow, Kubernetes and much of the open-source infrastructure that runs the internet. [5]

The change signals Google's commitment to open-source AI as a competitive strategy. By making its most capable models freely available under a permissive license, Google is betting that developer adoption will drive ecosystem growth — and that ecosystem growth will benefit Google's cloud and developer tools businesses. [6]

The Competition

Gemma 4 enters a crowded field. Meta's Llama family dominates the open-model benchmark charts. Mistral and Qwen are strong competitors. The distinction Gemma 4 offers is the combination of Google's research pedigree, the Apache 2.0 license and the 256K context window. [7]

The 256K context is notable. It allows the model to process documents of roughly 200,000 words — an entire book, a large codebase, or months of conversation history — in a single prompt. Most proprietary models cap at 128K. Gemma 4 doubles that. [8]

The Impact

Open-source models are changing the economics of AI development. Companies that cannot afford API access to proprietary models can run Gemma 4 on their own hardware. Researchers can fine-tune the models for specific domains without licensing restrictions. Startups can build products on top of Gemma 4 without negotiating with Google. [9]

The Apache 2.0 license makes this possible. It is the difference between "open weights" and "open source." Gemma 4 is the latter. [10]

-- DAVID CHEN, Beijing