Google has released EmbeddingGemma 2, an embedding model that maps text, code, images, video, and audio into a single vector space, so a device can retrieve a photo or a video clip from a text description without contacting a server. The family totals 740M parameters: 270M for text, 170M for images, and 300M for audio, and each module can be attached on its own. In an optimized run on a Pixel 11 Pro, the full model uses roughly 567 MB of active RAM, while the text-only path stays near 191 MB. The weights ship under Apache 2.0, which permits commercial use, and the intended workload is local search over documents and media where data cannot leave the device.

EmbeddingGemma

Related: Google Releases Gemma 4 12B, Google Unveils Coral Board: A RISC-V SBC for On-Device Gemma 3