Google Gemma 4: The Autonomous Multilingual Mobile AI

A Local and Autonomous AI with Gemma 4

Google recently unveiled Gemma 4, an open-source artificial intelligence model that operates entirely on-device. This model is capable of processing text, images, and audio without the need to transfer data to external servers. With integrated agent capabilities, Gemma 4 can utilize tools like Wikipedia, interactive maps, and QR code generators, all while remaining autonomous.

The smartphone variants, E2B and E4B, are designed to run on devices with 6 and 8 GB of RAM, respectively. According to Google, these models offer processing speeds up to four times faster than the previous generation. These innovations foreshadow the upcoming Gemini Nano 4 intended for Android devices.

Gemma 4 is built on the same research as Google's proprietary model, Gemini 3. Since the launch of the first generation, the Gemma family has accumulated over 400 million downloads. All models in this family handle text, images, and audio in more than 140 languages.

All these models are distributed under the Apache 2.0 license, which is particularly favorable for businesses. This license allows developers to create and share custom skills via GitHub. Additionally, the free app Google AI Edge Gallery is available for Android and iOS users.

Models Tailored to Every Need

The latest version of Gemma 4 comes in four variants, each tailored to specific needs. The E2B and E4B models are optimized for smartphones, with the "E" standing for "effective parameters," meaning the number of parameters that are actually active during inference. The E2B model occupies about 1.3 GB of memory on the device, while the E4B model requires about 2.5 GB.

For more demanding needs, the 26B and 31B variants are intended for servers and high-performance hardware. The 26B model uses a mixture of experts architecture with 128 experts, meaning that only 3.8 billion parameters are active at any given time. The dense 31B model offers an impressive context window of up to 256,000 tokens.

Google has collaborated with Arm and Qualcomm to optimize these models for current mobile chips. According to Google, Gemma 4 on Android operates up to four times faster than the previous generation while reducing battery consumption by 60%. Arm's benchmarks even indicate average speed gains of 5.5 times for devices equipped with recent Arm chips using the SME2 instruction set.

Agent Skills for Optimized Use

The Gemma 4 application requires Android 12 or iOS 17 to function. The E2B and E4B models differ in their RAM requirements: E2B uses about 1.3 GB and runs on devices with 6 GB of RAM, while E4B requires about 2.5 GB of model memory and at least 8 GB of RAM.

Agent skills can be activated and managed individually, allowing Gemma 4 to generate a QR code directly on the device using a JavaScript skill. In addition to basic chat functions, image recognition, and audio transcription, the application offers "agent skills" such as Wikipedia search, interactive maps, automatic summaries, and flashcards.

Gemma 4 is also capable of describing photos, transforming voice inputs into diagrams and visualizations, and collaborating with other local models for tasks like text-to-speech or image generation. Google demonstrated these capabilities with a showcase skill that describes and plays animal calls.

Image recognition has been significantly improved, especially for OCR tasks that extract text from images, diagrams, or handwritten notes. The model also better handles temporal information, which is crucial for calendars, reminders, and alarms.

Gemma 4 can detect the intent behind user input and automatically activate the appropriate skill, such as a mood tracker with a history graph.

While these features may not be revolutionary compared to existing cloud services, the fact that a demo application can run a purely local model on a phone and use these tools autonomously is remarkable. Developers can create and share custom skills via GitHub, and although the integrated tools require an internet connection, the model itself operates locally, and conversations are never saved.

Towards the Future with Gemini Nano

Google indicates that Gemma 4 E2B and E4B serve as a foundation for the future Gemini Nano 4, the next generation of on-device models for the Android system. The code developed for Gemma 4 today will be compatible with Gemini Nano 4 upon its release on new flagship devices later this year. Gemini Nano is already operational on over 140 million Android devices, powering features like Smart Replies and audio summaries.

In December, Google introduced FunctionGemma, a small local model with only 270 million parameters, capable of directing commands to other applications on the phone. It translates natural language into structured function calls: turning on the flashlight, creating contacts, sending emails, adding calendar entries, displaying locations on a map, or opening Wi-Fi settings.

The strategic importance of on-device AI was highlighted earlier this year with a billion-dollar agreement between Apple and Google. Since January, it has been known that Apple's next generation of Foundation Models will be built on Google's Gemini technology, supporting a complete overhaul of Siri planned for 2025.