Google Opens Gemini Embedding 2 Preview for Multimodal Retrieval

Google AI Developers said on March 10, 2026 that Gemini Embedding 2 is now in preview via the Gemini API and Vertex AI. The company described it as its most capable embedding model and the first fully multimodal embedding model built on the Gemini architecture.

In its official blog post, Google said Gemini Embedding 2 maps text, images, video, audio, and documents into a single representation space. The launch also includes support for 100+ languages and multiple output dimensions, which gives developers more flexibility when they need to trade off recall, latency, and storage cost.

This is a material release for retrieval and ranking workloads. Teams building search, RAG, recommendation, moderation, or cross-media discovery systems often end up stitching together separate embedding stacks for different modalities. A natively multimodal model reduces that fragmentation and can simplify indexing pipelines across product surfaces.

Why it matters

• One embedding space across text, image, video, audio, and PDF workflows can simplify retrieval architecture.
• Preview availability in both the Gemini API and Vertex AI lowers friction for Google Cloud users.
• Flexible dimensions give teams a practical knob for cost and performance tuning.

Because the release is still in preview, pricing, limits, and benchmark positioning can evolve. But the direction is clear: Google wants Gemini to compete not only as a generation model family, but as a default platform layer for multimodal retrieval. That could matter for enterprise search and agent systems that increasingly span more than plain text.

Primary sources: Google AI Developers on X and Google Blog.