EmbeddingGemma 2 brings multimodal embeddings on-device

Google released EmbeddingGemma 2, an embedding model small enough to run on a phone that now handles text, images, audio, video, and code in a single shared vector space. The full model is 740 million parameters, but it is built to be taken apart: a 270M text-only core with optional 170M vision and 300M audio encoders, so you load only the modalities you need. On a Pixel 11 Pro the text-only path uses about 191MB of RAM with quantization, and the full multimodal setup about 567MB.

The point of a unified embedding space is cross-modal search without a server: you can match a text query against images or a clip of audio directly on the device. The context window grew to 8K tokens, four times the first version, enough for 29 images or roughly five and a half minutes of audio in one pass. On the MTEB Code benchmark it scores 78.68, up nearly ten points from the first EmbeddingGemma. Matryoshka representation learning lets you truncate the 768-dimension vectors to 512, 256, or 128, cutting storage by up to six times when you can accept some quality loss. It ships under Apache 2.0 on Hugging Face and Kaggle.

Why it matters

If you build search or retrieval into mobile or edge apps, this lets you index photos, voice notes, and text with one model and no network round trip, which helps with privacy and offline use. The Apache 2.0 license means you can ship it in commercial products.