Google has released Gemma 4, a family of four open-weight artificial intelligence models licensed under the permissive Apache 2.0 license, marking a significant strategic shift in the global open-source AI race. The launch, announced on April 2, 2026, by Google DeepMind CEO Demis Hassabis, positions the new models as the strongest American contenders against a field recently dominated by Chinese alternatives like DeepSeek and Qwen.
The Gemma 4 lineup consists of four variants designed for different hardware: the Effective 2B and Effective 4B models for phones and edge devices; a 26B Mixture of Experts (MoE) model optimized for speed and low latency; and a flagship 31B Dense model built for raw performance. According to Arena AI's text leaderboard, the 31B model currently ranks third globally among all open models, with the 26B MoE sitting in sixth place. Google claims these models can outperform competitors 20 times their size.
The release is a direct response to the shifting landscape of open-source AI. Over the past year, Chinese models from companies like Alibaba (Qwen), DeepSeek, Minimax, and GLM have surged in popularity, growing from roughly 1.2% of global open-model usage in late 2024 to about 30% by the end of 2025. This rise coincided with a decline in the relevance of previous Western standards like Meta's Llama, whose restrictive license and slipping performance created an opening.
Gemma 4 is built on the same research as Google's proprietary Gemini 3 model. Key features include advanced reasoning for multi-step logic, native support for agentic workflows with function calling and structured JSON output, and high-quality offline code generation. The models support over 140 languages and feature expanded context windows—128K tokens for edge variants and 256K for the larger models.
A critical change is the adoption of the Apache 2.0 license, a departure from the custom, more restrictive licenses used for previous Gemma versions. This move eliminates legal ambiguity for commercial use, allowing developers to freely modify, redistribute, and commercialize the models. Hugging Face co-founder Clement Delangue praised the decision, stating, "Local AI is having its moment."
The models are available immediately. The 31B and 26B variants can be accessed via Google AI Studio, while the E2B and E4B edge models are in the Google AI Edge Gallery. Weights are also available on Hugging Face, Kaggle, and Ollama.
