Google Releases Gemma 4 Under Apache 2.0, Unlocking Fully Open-Source AI

Google released Gemma 4 on Thursday as its first fully open-source AI model, distributed under the Apache 2.0 license and marking a strategic departure from the company's previous approach to model distribution.

Unlike earlier Gemma iterations, which were open-weight but bound by Google's terms of service, Gemma 4 permits unrestricted modification, commercial use, and redistribution with only attribution required. The model can run locally on billions of Android devices, laptop GPUs, and edge hardware including Raspberry Pi units, eliminating cloud dependencies and associated costs.

The licensing shift addresses enterprise requirements that previous open-weight releases could not satisfy. Healthcare providers bound by data sovereignty rules, manufacturers operating in low-connectivity environments, and organizations seeking to avoid recurring cloud fees can now deploy capable AI without sending data off-premises. Google AI Studio, Hugging Face, Kaggle, and Ollama are distributing the model.

(The release arrives as Microsoft announced three proprietary enterprise models on its Foundry platform, signaling divergent strategies among hyperscalers over open versus closed AI development.)

Google's move intensifies competition in the open-source AI arena, where Meta has championed permissive licensing through its Llama series and Chinese developers have captured significant market share with models like DeepSeek. The Apache 2.0 designation places Gemma 4 among the most permissively licensed frontier models from a major US technology company, potentially reshaping developer expectations around model access and control.

The decision to fully open-source a capable multimodal model reflects broader industry tension between proprietary advantage and ecosystem growth. While OpenAI and Anthropic maintain closed development, Google appears to be wagering that widespread adoption and community contribution will yield strategic returns that outweigh the risks of unrestricted distribution.