Introducing Gemma 4: Google’s New Open AI Models for On-Device Intelligence
On April 2, 2026, Google DeepMind unveiled Gemma 4, a cutting-edge family of open-weight AI models. Released under the lenient Apache 2.0 license, this marks a significant stride towards integrating robust AI capabilities directly into consumer devices and enterprise edge computing.
The Gemma 4 family comprises four model sizes, each optimized for a variety of use cases ranging from smartphones to data centers. These models boast advanced reasoning capabilities, native support for over 140 languages, and multimodal inputs including text, images, and audio. Additionally, they offer native function calling. Designed to run efficiently on local hardware, these models eliminate the need for constant cloud connectivity, thereby addressing data privacy concerns.
Key innovations of Gemma 4 include a 26 billion-parameter mixture of experts (MoE) model. This model activates only 3.8 billion parameters during inference, ensuring faster response times without compromising on quality. With a 256,000-token context window, these models are ideally suited for local code assistants and complex reasoning tasks. Google reports that Gemma 4 on Android devices performs up to 4x faster than its predecessors and consumes up to 60% less battery power.
The launch of Gemma 4 signifies Google’s answer to the surge of competitive open-weight models from Chinese companies such as Moonshot AI and Alibaba. By providing a domestically-developed alternative under a permissive license, Google aims to attract enterprise customers seeking powerful AI capabilities without the worry of data being used for future model training. Gemma 4 is immediately available through Google AI Studio, Hugging Face, Kaggle, and other popular platforms, with day-one support for over a dozen inference frameworks.
Source: Google Developers Blog