Google DeepMind launched EmbeddingGemma 2, the successor to its on-device text-embedding model, which the company says has passed 20 million downloads. The new version "expand[s] beyond text to unify code, images, video, and audio in a shared embedding space," built on the Gemma 4 architecture and released under an Apache 2.0 license. At 740 million parameters it is designed for on-device inference, combining modular encoders — 270M text, 170M vision and 300M audio. Matryoshka truncation can shrink embeddings from 768 dimensions to 512, 256 or 128, for up to a 6x storage reduction, while context grows to 8K tokens, four times the prior generation. Google cites roughly 191MB of active RAM for the text-only path and about 567MB for full multimodal on a Pixel 11 Pro, and reports a 9.92-point gain on MTEB Code (68.76 to 78.68).
Our takeOur take: The multimodal unification is the more consequential part of this release — indexing photos, clips and audio locally is a privacy win, not just a latency one. A sub-billion-parameter model under a permissive Apache 2.0 license also keeps this within reach for smaller developers who can't call a frontier API for every lookup. The open question is quality: a single shared embedding space often trades some per-modality accuracy for convenience, so the MTEB Code gain will need independent replication.