Google Unveils New AI Model for Smartphones That Works Offline

Google has unveiled a new open-source model called EmbeddingGemma 2 to expand AI capabilities on smartphones and other devices. With just 740 million parameters, the model can search text, images, video, audio and software code without an internet connection. According to the company, the new model also outperforms some competing models twice its size in performance tests.

EmbeddingGemma 2's main advantage is that it connects different types of data within a single system. By converting text, images, video and audio into numerical vectors, the model determines their semantic similarity. For example, a user can enter text to find a particular topic in hours of audio recordings or search for video footage matching a phrase in an audio recording.

The first EmbeddingGemma, introduced by Google last year, worked only with text. The new version is based on the Gemma 4 architecture and supports software code, images, video and audio in addition to text. The model is available under the Apache 2.0 license.

Although EmbeddingGemma 2 has a total of 740 million parameters, using all of the model's capabilities is not required for every task. It is possible to use sections with 270 million parameters for text operations, 170 million for images and 300 million for audio data. This approach conserves the device's computing resources.

Another advantage of the model is optimized memory usage. Thanks to Matryoshka Representation Learning (MRL) technology, the generated 768-dimensional vectors can be reduced to 512, 256 or even 128 dimensions. According to Google, this method can reduce memory and storage requirements in local vector databases by up to six times.

EmbeddingGemma 2 also has a larger context window than the previous version. The model's 8,000-token context window is four times larger than that of the previous version. This makes it possible to process approximately 5.5 minutes of audio, 29 images or 58 video frames directly on the device.

Google has optimized the new model specifically to run locally on smartphones. According to the company, EmbeddingGemma 2 requires approximately 191 MB of active RAM on the Google Pixel 11 Pro when performing text-only operations. When other data types are also used, this figure rises to approximately 567 MB.

The model can also run directly in a browser through WebGPU technology. Based on figures provided by Google, queries can be processed within 20–70 milliseconds under these conditions. This eliminates the need to send data to remote servers or use an API key.

This approach is particularly important for personal data privacy. Because data is processed on the device itself, users' files and queries can be indexed for search without being transmitted to external servers. The lack of a need for an internet connection also allows AI-powered search systems to operate in environments where network access is limited.

EmbeddingGemma 2's model weights are available on the Hugging Face and Kaggle platforms. Support for the Gemini Enterprise Agent Platform's Model Garden service is also planned for the future. Versions optimized for use on devices are available on the LiteRT Community's Hugging Face page.

The new model is regarded as an important step toward enabling AI to operate on smartphones without relying solely on cloud services. However, EmbeddingGemma 2 is not a general-purpose chatbot; it is a specialized model designed to identify semantic relationships among different types of data and improve search capabilities.