A 740M Model Now Does What Five Specialist Models Did
One small open embedder handles text, code, image, audio and video. That changes on-device search.
Google released EmbeddingGemma 2 on October 6 as an open 740-million-parameter model under Apache 2.0 that embeds text, code, images, audio and video in one representation. The context window went to 8,000 tokens, four times the prior version.
For most teams this kills a whole category of plumbing. You used to run a separate embedder per data type and stitch the results together. Now one model covers the lot.
It's small enough to run offline on a phone -- Google cited roughly 191MB of active RAM for text-only use. That means search and retrieval can live on the device instead of calling a server for every query.
I'd still benchmark it on your own corpus before trusting it for legal or medical retrieval. Open and small is a great default, not a guarantee your domain fits.