Google launches Gemini Embedding 2 for unified text, image, audio, video, and document search

Google AI Studio highlighted Gemini Embedding 2 in a March 12, 2026 post on X, saying the model can bring text, images, audio, video, and documents into a single vector space. Google’s March 10 product post adds the deeper technical framing: Gemini Embedding 2 is Google’s first fully multimodal embedding model built on the Gemini architecture, and it is available in public preview through the Gemini API and Vertex AI.

That matters because embeddings sit underneath many production AI systems. Search, retrieval, recommendation, clustering, sentiment analysis, and Retrieval-Augmented Generation all depend on turning content into representations that can be compared efficiently. Historically, teams often handled each modality with separate pipelines. Google’s claim here is that Gemini Embedding 2 reduces that fragmentation by mapping text, images, videos, audio, and documents into one unified embedding space.

In its blog post, Google says the model captures semantic intent across more than 100 languages and is designed for multimodal retrieval and classification. It also points developers to the Gemini API, Vertex AI, and an interactive semantic search demo. Google further says the model sets a new performance standard for multimodal depth and shows strong speech capabilities. Those benchmark and performance statements are Google’s own claims, and should be read that way.

The most concrete user example in the official material comes from Paramount Skydance. According to Google’s post, the company used Gemini Embedding 2 to let text queries retrieve matching video assets, including untranscribed micro-expressions, and reported a text-to-video Recall@1 rate of 85.3%. If that result generalizes, it would be a meaningful sign that unified multimodal embeddings are moving from research promise to operational media workflows.

For developers, the strategic takeaway is straightforward. A single embedding space can simplify multimodal search and RAG stacks, reduce glue code between modality-specific systems, and make it easier to build applications that search across mixed corpora instead of just text. The launch does not remove the need to test quality on domain data, but it does show Google pushing multimodal retrieval closer to default infrastructure.

Primary sources: Google AI Studio on X and Google’s product post.