Google explains how it trained EmbeddingGemma 2, a 744M-parameter model that searches across text, images, audio and video
Cross-media search often turns speech and images into text, losing detail, or needs multibillion-parameter models too big for phones. Google's EmbeddingGemma team detailed EmbeddingGemma 2, a 744M-parameter model turning text, images, audio and video into comparable number vectors. Compressed, it needs about 191 MB on a phone for text. Its general English score dips slightly from its predecessor.
Why it matters
Developers can use this Apache 2.0 open model to search across text, images, audio and video right on a phone, which suits apps that need privacy or offline use.
Sources2
Written by dalao.ai: the title, summary and why it matters are our own words. The original reports are the record; their publishers keep authorship.