What's this about?
The 740M model maps text, code, images, audio and video into one embedding space. Developers can run only the 270M text component or add vision and audio modules as needed.
What exactly is an embedding?
It is a compact numerical coordinate for meaning. Items with related content land near one another, allowing search across different formats without matching exact words.
What's the community saying?
DeepMind's announcement drew 3,150 likes; Sundar Pichai's drew 5,759.
- Offline multimodal search offers privacy and avoids cloud latency.
- Modular encoders and shorter vectors reduce RAM and storage on phones.
- Developers still need real-device quality and battery benchmarks beyond vendor leaderboards.