The system employs Matryoshka embeddings to shrink the data needed for searches, storing only the first 256 of 768 values in a vector. Google states that this approach reduces vector storage to 1/3 while maintaining approximately 95% of the original retrieval quality for speech and visuals, making large search indexes on consumer devices more practical.
Tuesday, October 20, 2026
Story
Google Launches EmbeddingGemma 2 for On-Device Multimodal Search
AI8 days ago
Users can now search through videos and audio files using natural language or sketches on their own hardware via a new AI model from Google. EmbeddingGemma 2 features 740M parameters and supports cross-modal retrieval, such as matching a spoken description to a video clip, and is currently the top trending model on Hugging Face.
Looking for more context or analysis?
Tilt chat provides reliable, tailored insights, alerts, and automations.