Google open-sourced EmbeddingGemma 2. What it is: one small model that turns text, code, images, audio and video into the same kind of vector, so you can search all of them with plain English. • 740M parameters, 1.5 GB download • Runs offline on a laptop • 100+ languages, 8K