Google released EmbeddingGemma 2 in October 2026: an open, lightweight embedding model that turns text, code, images, video and audio into vectors, designed to run on the device itself. It has about 740 million parameters (around 270 million if you only need text), an 8,000-token context window, flexible output sizes from 768 down to 128 dimensions, and an Apache 2.0 licence. Third-party coverage reports it runs in a few hundred megabytes of memory on a phone.
What embedding models do
An embedding model converts content into a list of numbers that captures its meaning. Similar meanings produce similar vectors. That powers:
- Semantic search: "cotton kurta for summer" finds the right products even without exact keywords.
- Retrieval for AI chatbots (RAG): finding the right passages from your documents to answer a question.
- Recommendations and deduplication: "more like this", and spotting near-duplicate listings.
- Classification and clustering of tickets, reviews or leads.
Why on-device matters
- Privacy: personal photos, notes and messages never leave the device.
- Cost: no per-request API fee.
- Speed and offline use: no network round trip.
EmbeddingGemma 2's multimodal support means one model can search across text, images and audio in the same space. That's useful for photo search, voice notes and product catalogues.
Where it fits in your product
- Mobile apps: private on-device search over the user's own content.
- Browsers and edge: smaller vectors and quantised models make in-browser search feasible.
- Servers: an open model you can self-host for catalogue and document search without per-call fees.
Flexible output sizes are a practical win: smaller vectors mean cheaper storage and faster search in a vector database like pgvector, at a small cost in accuracy.
AI search for your website or app
At NextGen Fusion, our AI development company in Lucknow, we build semantic search and AI assistants for e-commerce catalogues and document libraries, choosing on-device, self-hosted or API models based on privacy and cost.
