Google releases EmbeddingGemma 2, a light open model

Google DeepMind has introduced EmbeddingGemma 2, the second generation of its open, lightweight embedding model, which is now multimodal and can work with more than just text. Because the model is open, companies get more freedom in deciding where and how to run it.

An embedding turns material into a set of numbers so that texts close in meaning end up next to each other. That is the basis of search by meaning: before answering, an agent pulls the relevant passages from an archive instead of guessing. Being lightweight matters because the model runs on every request, and multimodality extends this kind of search beyond typed text, which is significant for companies where half the knowledge sits in scans and photos.

Check what form your company's knowledge actually takes: how much of it lives in scans, photos of delivery notes and chat threads rather than tidy spreadsheets. Primary documents from photos, bank statements and document reconciliation are handled by an agent in 1Cifer, so nobody retypes a delivery note from a messenger. Start with one document flow and estimate how many hours a week go into manual searching.