AI

Google Launches EmbeddingGemma 2 for On-Device Multimodal Search

Monday, October 12, 2026Read Original

Details

  • Google introduced EmbeddingGemma 2, an open multimodal embedding model for searching text, images, video, and audio on consumer devices.
  • Google AI Edge enables developers to run the model locally, reducing reliance on cloud processing and supporting privacy-focused applications.
  • The 740-million-parameter model maps multiple media types into a shared 768-dimensional vector space, allowing searches such as finding video clips with a text query or voice memo.
  • Its modular design lets developers load only the encoders they need, with configurations ranging from text and code to full multimodal operation.
  • MediaPipe Tasks handles image resizing, tokenization, and search ranking, reducing the custom code required for cross-modal retrieval.
  • Google says the full model can use about 567MB of active RAM on a Pixel 11 Pro, while text-only operation can use about 191MB.
  • EmbeddingGemma 2 is built on the Gemma 4 architecture and released under the commercially permissive Apache 2.0 license.
  • The announcement extends earlier on-device embedding workflows by combining vision, audio, video, and text in one compatible model rather than separate modality-specific pipelines.

Impact

EmbeddingGemma 2 strengthens Google’s edge-AI strategy by moving multimodal search from cloud services onto phones and laptops. Its open license, modular architecture, and sub-gigabyte footprint could lower the engineering and infrastructure costs of local media retrieval, while improving privacy for sensitive files. The approach also increases competitive pressure on providers whose multimodal search depends primarily on hosted APIs, although real-world adoption will depend on device performance, tooling support, and retrieval quality across different media collections.

Rift Dispatch
Google Launches EmbeddingGemma 2 for On-Device Multimodal Search | RiftLab