What changed

Google released EmbeddingGemma 2 under Apache 2.0. The modular model maps text, code, images, video and audio into one 768-dimensional space, scaling from a 270M text-and-code path to a 740M multimodal path; Google reports about 191 MB and 567 MB active RAM respectively on a Pixel 11 Pro.

Why it matters

A shared compact embedding space can replace separate captioning, transcription and retrieval stages in private local search. The open weights and modular encoders make memory, modality and retrieval tradeoffs directly testable on target devices.

What to do

Pin the model revision and build a mixed text, image and audio retrieval set with privacy-safe data. Measure recall, cross-modal false matches, indexing time, active and peak memory and device latency against the current separate-model pipeline.

Caveat

Google's best-in-class and memory figures are vendor evidence tied to its evaluation and Pixel configuration. Open weights do not guarantee application-specific retrieval quality, and full multimodal use needs more memory than the text-only module.