Trusence Every claim has a source
Last updated 17 November 2026 Search Türkçe
Research

Google's EmbeddingGemma 2 adds images, audio and video to its open embedding model

A single shared embedding space now covers text and other media, with the technical report as the reference for exact results.

Google has released EmbeddingGemma 2, an update to its text-only open embedding model that adds native multimodal support. The model maps text, images, audio and video, or any mix of them, into one shared embedding space while using a more compact model backbone. A technical report now describes its training strategy, evaluation method and experimental results, and the model builds on techniques from Gemma 4 and Gemini Embedding 2. The report is the main source for the specific figures, so developers who need exact numbers should read it directly.

Why it matters

Embedding models turn content into vectors that search and retrieval systems compare, so one space covering several media types means text, images, audio and video can be matched against each other rather than handled by separate models. The update also moves an open model toward the multimodal capability already offered in Google's Gemini line, which matters for anyone following how open and proprietary models are converging. The model is described as using a more compact backbone, and the published technical report now sets out its training and evaluation, giving outside observers a documented basis for judging its results.

Sources · 1