Google has introduced its latest open AI model, Gemma 4. Differing from other cutting-edge models like Gemini or OpenAI’s GPT 5.4, Gemma 4 is an open-source model that can be operated locally on various devices, including Android smartphones. Gemma 4 provides developers with the flexibility of open models while retaining the same functionality as proprietary models.
Demis Hassabis, CEO of Google DeepMind, unveiled the release of Gemma 4, claiming that the four sizes of Gemma 4 are “the best open models globally for their respective sizes.” These models are open-source, allowing anyone to download Gemma 4 and run it locally on their device for free. This approach enhances privacy as the locally-run model does not share data with Google or any third party.
Sriram Krishnan, a White House policy advisor, emphasized the importance of open-source models for the Western world, stating that they play a crucial role in maintaining a competitive edge.
Gemma 4 comprises four purpose-built models for advanced reasoning and workflows: Effective 2 Billion (E2B) and Effective 4 Billion (E4B) for mobile and edge devices, 26 Billion Mixture of Experts (MoE) focusing on low latency, and 31 Billion Dense, offering high raw performance. The 31B model ranks third globally among open models on the Arena AI leaderboard, surpassing models 20 times its size.
Google highlights that Gemma 4 introduces significant advancements in capabilities such as multi-step planning and complex logic. These models also support workflows that allow users to delegate tasks to AI agents, enabling the creation of a team of AI agents to handle various tasks.
Gemma 4 is designed to run on a wide range of hardware, from high-performance GPUs in large computers to regular Android smartphones. It serves as the base model for the next generation of Gemini Nano, the foundational AI model on Android, potentially enabling billions of smartphones to run the AI model locally without an internet connection.
Since its initial launch, Gemma has been downloaded over 400 million times by developers, fostering a community ecosystem with over 100,000 variants. The models are released under an Apache 2.0 license, granting developers complete flexibility to use, modify, and deploy them.
Gemma 4 also offers high-quality offline code generation, local AI coding assistance, image and video processing capabilities, and support for multiple languages. Users can access Gemma 4 through platforms like Google AI Studio, Hugging Face, Kaggle, and Ollama.
Overall, Gemma 4 represents a significant advancement in AI technology, catering to a wide range of developers and users seeking cutting-edge AI capabilities.