Google Releases Gemma 4, Its Most Capable Open Model

Google DeepMind releases Gemma 4 under an Apache 2.0 license, intensifying the open-source AI arms race.

Google DeepMind today released Gemma 4, a family of open-source large language models built on the same research underpinning Gemini 3. CEO Demis Hassabis announced the launch, which marks a significant strategic shift: for the first time, Google's open models ship under the commercially permissive Apache 2.0 license, granting developers full control over deployment, modification, and redistribution.

The release spans four model sizes. The flagship 31B Dense variant ranks third on the Arena AI text leaderboard among all open models. A 26B Mixture of Experts model sits sixth. Two smaller variants, the E4B and E2B, were co-developed with the Pixel team, Qualcomm, and MediaTek to run on phones, Raspberry Pi boards, and Jetson Nano devices with near-zero latency. Google claims the edge models "outcompete models 20x their size."

All four models process video, images, and text natively, with the smaller variants adding audio input for speech recognition. Context windows reach 256K tokens for the larger models and 128K for the edge variants, enabling full-document and repository-scale prompts. The models support native function-calling, structured JSON output, and multi-step agentic workflows out of the box, positioning Gemma 4 directly against Meta's Llama and Mistral in the enterprise open-source race.

The Apache 2.0 licensing decision may prove more consequential than the benchmarks. Previous Gemma releases carried more restrictive terms. The shift to full open-source removes a key barrier for enterprise adoption and signals Google is betting that ecosystem dominance, not licensing revenue, wins the open-weight AI market.

Weights are available now on Hugging Face, Kaggle, and Ollama.