Google DeepMind has officially unveiled Gemma 4, the fourth generation of its open model family. Representing the most comprehensive release in the Gemma lineage, Gemma 4 brings the research and technology behind Google's proprietary Gemini 3 models to the open-source community. In a significant departure from previous releases, the entire model family is now available under the industry-standard Apache 2.0 license, granting developers full freedom to use, modify, and distribute the models for any purpose, including commercial applications.
The Gemma 4 family comprises four distinct model sizes. The E2B and E4B models, designed for edge and mobile deployment, feature 2 billion and 4 billion effective parameters respectively. For workstations and servers, the 26B A4B Mixture-of-Experts and 31B Dense models deliver substantially greater capability. According to Google's benchmark results, the 31B model secured the third position on Arena AI's text leaderboard, outperforming models 20 times its size. The smaller models support 128K context windows, while the larger variants extend to 256K tokens.
All models handle text and image inputs, while the E2B and E4B models additionally support audio processing. This multimodal capability enables tasks such as optical character recognition, speech recognition, video analysis, and object detection to be performed entirely on-device without sending data to the cloud. Gemma 4 also supports over 140 languages and comes equipped with agentic features including native tool use and multi-step planning, making it suitable for autonomous agent workflows.
The licensing shift marks a pivotal moment for the Gemma ecosystem. Previous versions operated under Google's proprietary terms of use, which imposed restrictions on redistribution and certain use cases — making them "open" but not truly "open source." With Apache 2.0, those limitations have been entirely removed. Since its initial launch, Gemma models have been downloaded over 400 million times, spawning more than 100,000 community-built variants. The transition to a fully open-source license is expected to accelerate this growth significantly. The models are now available through Hugging Face, Kaggle, Ollama, and Google AI Studio.