Gemma 4 Is Here: Google's Most Capable Open Models Yet Run Offline on Smartphones and Consumer GPUs

The tech giant has now launched its most advanced open-source AI model, named Gemma 4. This AI model is intended for use on a range of devices, from data center workstations to smartphones. With its first release, the Gemma AI model has been downloaded more than 400 million times and has resulted in a community of more than 100,000 variants of the AI model itself.

The CEO of Google, Sundar Pichai, is very enthusiastic about the launch of the Gemma 4 AI model. According to him, it contains a ton of intelligence despite its size. Demis Hassabis, a co-founder of DeepMind, claims it to be "the best open models in the world." The AI model is available in four different versions: a high-performance version named 31B, a fast version named 26B, and two lightweight versions named 2B and 4B, which can be used on your smartphone or tablet itself. The best part of this AI model is that it is completely free to use, tweak, and work on.

Four Gemma 4 models, one goal

Google is releasing Gemma 4 in four different sizes, which can range from small mobile devices to high-end developer computers:

• E2B (Effective 2 billion parameters)—Optimized for use in phones and IoT devices

• E4B  (Effective 4 billion parameters)—Optimized for use in phones and IoT devices

• 26B Mixture of Experts (MoE)—A mid-range powerhouse

• 31B Dense—The flagship, currently ranked #3 in the world for all open AI models according to the industry-standard Arena Ai leaderboard

That last one is particularly noteworthy. The 31B Dense model reportedly outperforms competitors 20 times its own size.

What Gemma 4 is capable of

According to Google, Gemma 4 is capable of a number of things, with the most significant features being:

Advanced reasoning: The AI is capable of handling complex planning and logic.

Agentic workflows: The AI is capable of native function calls, structured data output, and system instructions, which makes it possible for developers to create AI agents that can work with external tools and systems.

Code generation: The AI is capable of being run completely offline on a developer’s local machine, which makes it possible for developers to create a private AI coding assistant.

Vision and audio: The four models are capable of handling images and video natively. The two edge models are capable of handling audio input for speech recognition.

Long context windows: The edge models are capable of handling 128,000 tokens in a single prompt. The larger models are capable of handling 256,000 tokens.

140+ languages: The Gemma 4 AI has been trained natively on more than 140 languages, which makes it one of the most inclusive AIs available.

Gemma models for smartphones

Another remarkable aspect of Gemma 4 is that it is still very powerful in spite of its small size. The E2B and E4B models were created in partnership with Google Pixel, Qualcomm, and MediaTek, who have created billions of Android devices. The device can run entirely offline with zero latency for normal devices like smartphones, Raspberry Pi, and Nvidia Jetson. The device is based on the same research as Google's top-of-the-line Gemini 3, which is for premium AI in normal hardware.