Google ships EmbeddingGemma 2 for on-device search

What happened: Google DeepMind launched EmbeddingGemma 2, a 740 million parameter open embedding model released under the Apache 2.0 license. It maps text (including code), images, video and audio into one shared embedding space, so a single model can find a specific video clip from a voice memo or search hours of audio with a text query. It runs on-device, needing as little as roughly 191MB of active RAM for text-only weights and about 567MB for the full multimodal model on a Pixel 11 Pro, with an 8K token context window.

Context: It follows the first EmbeddingGemma, which passed 20 million downloads, and is built on the Gemma 4 architecture with modular text, vision and audio encoders. Matryoshka representation learning lets developers truncate output vectors from 768 dimensions down to 128 for up to six times less storage, and code retrieval improves by 9.92 points over the first version on MTEB Code.

Why it matters: Multimodal retrieval has usually meant chaining separate captioning, transcription and text embedding models. Collapsing that into one small local model cuts latency, memory and the amount of data that leaves the device, which is the whole point for private on-device search and RAG.

(Original source: Google (The Keyword))