Google DeepMind has released EmbeddingGemma 2, a 740-million-parameter open-weight model designed to bring multimodal AI search to resource-constrained devices such as smartphones. It maps text, images, video frames and audio into a shared vector space, enabling searches across media without a cloud connection, Hackster.io reports.
Google DeepMind Watch analysis
What happened
The model is intended for local tasks including media search and zero-shot classification. Hackster says it requires about 567 MB of active RAM and can be integrated into apps using MediaPipe Tasks and LiteRT. Its modular architecture and quantisation-aware training are designed to keep memory use low.
Read Hackster.io’s report on EmbeddingGemma 2.
Why it matters
A model that can match ideas across text, images, audio and video on a phone could make private, offline search more practical. It also gives app developers a route to add multimodal features without sending every query to a cloud service. That is a useful change in where AI can run, not just another model-size contest.
Our read
The appealing part is the combination: several media types, local operation and a memory footprint aimed at devices with limited resources. If the reported requirements hold up in real apps, developers could build search that works on a user’s own device, even when the network does not. The next test is whether the model’s speed and search quality are good enough for everyday use.
What to watch
- How quickly developers add EmbeddingGemma 2 to phone and edge applications.
- Whether local search performs well across different kinds of media.
- How its memory use and speed compare on actual devices.
Discussion spark: Would you trust a phone-based AI search tool more if it kept your photos and recordings on-device, even if its results were less capable than a cloud service?
Sources and evidence
- EmbeddingGemma 2 Puts Multimodal AI Search in Your Pocket – Hackster.io (16 October 2026, 19:53 UTC)
not affiliated with, endorsed by, or operated by Google or Google DeepMind