Google makes Gemini Embedding 2 generally available

Building multimodal search used to mean stringing together a fragile pipeline of separate models: one for text, another for images, and complex workarounds for audio and video.

Google just announced the general availability of Gemini Embedding 2, rolling it out across the Gemini API, Vertex AI, and the Gemini Enterprise Agent Platform.

One vector space for everything

Gemini Embedding 2 is natively multimodal, capable of mapping text, images, video, and audio into a unified embedding space. That eliminates the need to juggle fragmented data pipelines just to let users search video content with a text prompt.

During its preview phase, developers used the model to build advanced e-commerce discovery engines and video analysis tools. Google notes that this release introduces production-grade stability and latency optimizations for technology that already powers several of its internal services.

Why it matters

If you build multimodal search engines, media asset managers, or enterprise knowledge bases, this model simplifies your infrastructure overnight. Being able to reason across audio, visual, and textual data with a single endpoint makes multimodal RAG much more practical to run at scale.

While rivals focus heavily on multimodal generation, Google is quietly strengthening the foundational retrieval layer. Turning their battle-tested internal retrieval tech into an accessible API looks like a very smart play to anchor enterprise developers to their AI stack.