The new on‑device model handles text, code, images, video and audio, letting users perform tasks such as locating a video clip from a voice memo without sending data to the cloud.

Google unveiled EmbeddingGemma 2, a 740‑million‑parameter multimodal AI model that can run entirely on smartphones, expanding the company’s push for privacy‑first, on‑device intelligence.

What EmbeddingGemma 2 Can Do

The new model supports text, code, images, video, and audio, enabling a range of tasks such as searching for a specific video clip within a voice memo, generating code snippets from spoken commands, or extracting key frames from a video without ever uploading the data to Google’s servers.

Performance and Efficiency

Despite its size, EmbeddingGemma 2 is optimized for mobile CPUs and GPUs, delivering inference times comparable to earlier on‑device models while using less than half the battery consumption reported for its predecessor.

  • Runs on Android 13 and later devices
  • Supports offline operation for all supported modalities
  • Integrates with Google’s existing AI APIs for developers

Privacy Implications

By processing data locally, the model reduces the need to transmit potentially sensitive information to the cloud, aligning with Google’s broader privacy roadmap and offering users greater control over their personal data.

Developers can now embed EmbeddingGemma 2 into apps via the updated Google AI SDK, opening new possibilities for on‑device personalization and real‑time media analysis.

EmbeddingGemma 2 represents a significant step toward truly private AI experiences on mobile devices.

The model will be rolled out through Google Play Services later this month, with broader availability expected across a wide range of Android smartphones.

Read The Verge coverage of Google’s on‑device AI model launch.