Google has released EmbeddingGemma 2, a new version of its open embedding model that adds native multimodal understanding. The model maps text, images, audio, and video — or any combination of them — into a single embedding space, while compressing the model backbone into a more compact profile. The release builds on technology from Gemma 4 and Gemini Embedding 2, which Google says helps the model learn an expressive and unified vector space. Google has published the full technical report covering its training strategy, evaluation methodology, and experimental results, along with a five-minute explainer video showing how EmbeddingGemma 2 works and how it was developed.