Google announced its latest open AI model called Gemma 4. Unlike most Frontier models such as Gemini and OpenAI’s GPT 5.4, Gemma 4 is an open source model that can run locally on a wide range of devices, including Android smartphones. According to the company, Gemma 4 gives developers the freedom of an open model with the same level of functionality as a proprietary model.
Google DeepMind CEO Demis Hassabis announced its release in X. He claimed that the four sizes of Gemma 4 were “the best open models in the world in their respective sizes”. Here are the details:
Unlike Gemini, these models are open source. This means anyone can download Gemma 4 and run it locally on their device for free. The locally run model also provides more privacy, as it doesn’t share your data with Google or other third parties.
White House policy adviser Sriram Krishnan said the open source model is critical for Western countries. “The open source model is an important front for Western countries to lead, and this is a very important addition to this effort,” he said.
What is Gemma 4?
Gemma 4 consists of four sizes of purpose-built models for advanced inference and agent workflows. The Gemma 4 family includes four models designed for mobile and edge devices: Effective 2 Billion (E2B) and Effective 4 Billion (E4B). 26 billion experts (MoE) focused on low latency. 31 Billion Dense offers the best of raw performance.
The 31B model ranks third in the world among open models on the Arena AI leaderboard, outperforming models 20 times its size.
According to Google, Gemma 4 brings a huge leap forward in advanced features such as multi-step planning and complex logic. Gemma 4 models also support agent workflows that allow you to delegate tasks to AI agents. You can use Gemma 4 to build teams of AI agents to perform tasks like OpenClaw.
Gemma 4 on Android
Gemma 4 is designed to run on a variety of hardware, from large computers with high-performance GPUs to regular Android smartphones. According to Google, Gemini 4 is the base model for the next-generation Gemini Nano, the foundational AI model for Android.
This could potentially allow billions of smartphones to run AI models locally, without requiring an internet connection. AI models like ChatGPT and Gemini typically run via the cloud, and you can’t access them from your phone without an internet connection.
Google goes open source with Gemma 4
According to Google, since the launch of the first generation, developers have downloaded Gemma over 400 million times and built a community ecosystem with over 100,000 variations. The model is released under the Apache 2.0 license, giving developers complete flexibility in using, modifying, and deploying the model.
Please note that the previous Gemma model was not open source and only the open weights, i.e. the training dataset, was publicly available. And users still had to run those models according to Google’s terms.
What else can you do with Gemma 4?
Gemma 4 also offers high-quality offline code generation and allows you to use local AI coding assistants. All models can process images and video, but the smaller E2B and E4B also process audio input for voice recognition.
Gemma 4 supports longer context windows than previous generations. Edge models can handle up to 128,000 tokens and larger models can handle up to 256,000 tokens, allowing you to handle long documents and code repositories with a single prompt. Models are trained on over 140 languages.
Gemma 4 is accessible through Google AI Studio and can be downloaded from platforms such as Hugging Face, Kaggle, and Ollama.
– end
