EmbeddingGemma 2 brings multimodal search on-device

Transcript

The original narration, preserved as published. Relative dates refer to the episode’s coverage window.

Google just released EmbeddingGemma two: a small model that puts text, code, images, audio, and video into the same searchable space. The interesting part is where it runs: on your own device.

The full model has seven hundred forty million parameters. Its text, vision, and audio components are modular, so an app can load only the parts it needs. The weights are available now.

Picture finding a video clip using a voice memo, or searching hours of audio with a sentence. Those are cross-modal retrieval examples. The model returns vectors; a separate app retrieves and displays the results.

Google supports smaller output vectors, down from seven hundred sixty-eight dimensions to one hundred twenty-eight. That saves storage, but the model card shows accuracy trade-offs, especially for multimodal tasks. Test your own material.

This is an Apache two point oh release, with model weights and developer tools. Local inference can keep embedding generation on-device. Whether the entire application stays private depends on the rest of the pipeline.