Google Releases Gemini Embedding 2 — One Vector Space for Text, Images, Video, and Audio

Building a search system that handles text is straightforward enough. Adding images makes it harder. Throw in video and audio, and suddenly you’re maintaining three or four separate embedding pipelines, each with its own model, its own vector index, and its own set of headaches.

Google’s answer to this problem is Gemini Embedding 2, its first natively multimodal embedding model. Released on March 10, 2026, in public preview via the Gemini API and Vertex AI, the model maps text, images, video, audio, and PDF documents into a single unified vector space — no separate pipelines required.

The Problem With Multimodal Embeddings Until Now

Embeddings are the backbone of modern search, retrieval-augmented generation (RAG), recommendation systems, and classification tasks. But most embedding models only handle one modality. If you wanted to build a system where users could search a video library using a text query, you’d need to bridge multiple models, align their vector spaces, and hope the results made sense.

OpenAI’s text-embedding-3-large handles text well. CLIP-based models handle image-text pairs. But nothing on the market offered a single model that could natively process text, images, video, audio, and documents together — until now.

The practical cost of this fragmentation is real: separate infrastructure, separate indexing, separate maintenance, and degraded quality at the seams where modalities meet.

How Gemini Embedding 2 Tackles It

The core idea is simple: one model, one vector space, all modalities. You send in a text query, an image, a video clip, an audio file, or a PDF, and you get back a vector in the same embedding space. That means a text query can retrieve relevant video clips, an image can find similar audio segments, and cross-modal search works out of the box.

Here’s what the model can handle per request:

Text: Up to 8,192 tokens
Images: Up to 6 images (PNG or JPEG)
Video: Up to 128 seconds (MP4 or MOV)
Audio: Up to 80 seconds (MP3 or WAV)
PDFs: Up to 6 pages per request

The model also supports interleaved input — you can pass an image and text together in a single request, and the model captures the relationship between them rather than treating each modality in isolation.

Gemini Embedding 2 uses Matryoshka Representation Learning, a technique that lets developers truncate embedding dimensions without retraining. The default output is 3,072 dimensions, but you can scale down to 2,048, 1,536, 768, 512, 256, or even 128 dimensions depending on your latency and storage constraints. At 768 dimensions, the model still scores 67.99 on the MTEB benchmark — not far from the 68.17 it achieves at 1,536 dimensions.

Eight task types are explicitly optimized: semantic similarity, classification, clustering, document retrieval, query retrieval, code retrieval, question answering, and fact verification.

Where It Stands Against the Competition

The embedding model space has gotten crowded, so how does Gemini Embedding 2 compare?

vs. OpenAI text-embedding-3-large: Google’s model outperforms OpenAI’s flagship embedding model by roughly 6% on downstream task benchmarks, according to third-party evaluations. More importantly, OpenAI’s model is text-only — there’s no native support for image, video, or audio embeddings. If you need multimodal retrieval, you’d have to combine it with CLIP or another model.

vs. Cohere embed-v4.0: Cohere’s latest embedding model supports text and images but lacks native video and audio support. Gemini Embedding 2 covers more modalities in a single model.

vs. CLIP-based models: CLIP handles image-text pairs well but doesn’t support video, audio, or documents natively. It also wasn’t designed for tasks like code retrieval or fact verification.

The direct audio embedding capability is particularly noteworthy. Most systems require a transcription step (speech-to-text) before embedding audio content, which loses paralinguistic information like tone, pace, and speaker characteristics. Gemini Embedding 2 embeds audio directly, preserving that information.

What It Costs and How to Get Started

Pricing is straightforward:

Text embeddings: $0.20 per million tokens
Batch API: $0.10 per million tokens (50% discount for non-real-time workloads)
Image, video, and audio: Standard Gemini API media token rates

The model is available as gemini-embedding-2-preview through the Gemini API and Vertex AI. Integration support is already live for LangChain, LlamaIndex, Haystack, Weaviate, Qdrant, ChromaDB, and Vector Search.

One important caveat: embeddings from the previous text-only model (gemini-embedding-001) are not compatible. If you’re migrating a production system, you’ll need to re-embed your entire dataset. That’s a non-trivial cost for large-scale deployments, so plan accordingly.

Limitations to Keep in Mind

The model is still in public preview, which means API capacity may be limited and the model specification could change before general availability. A few other things worth noting:

Audio cap: 80 seconds per request is enough for short clips but won’t cover full podcast episodes or long-form audio without chunking
Video cap: 128 seconds covers most short-form content, but longer videos need to be split
Preview status: Production workloads should account for potential breaking changes
Re-embedding cost: Migrating from gemini-embedding-001 requires a full re-index, which can be expensive at scale
Language coverage: While 100+ languages are supported, performance may vary across less-resourced languages

FAQ

Is Gemini Embedding 2 free to use?
Text embeddings cost $0.20 per million tokens, with a 50% discount available through the batch API for non-real-time workloads. Image, video, and audio inputs follow standard Gemini API media pricing.

How does Gemini Embedding 2 compare to OpenAI’s embedding models?
Gemini Embedding 2 outperforms OpenAI’s text-embedding-3-large by approximately 6% on benchmark evaluations. The key differentiator is multimodal support — OpenAI’s model handles text only, while Google’s model natively processes text, images, video, audio, and PDFs in a single vector space.

Can I use Gemini Embedding 2 for RAG applications?
Yes. The model is designed for RAG workflows, supporting semantic search and document retrieval across multiple modalities. You can build systems where text queries retrieve relevant images, video clips, or audio segments alongside text documents.

What’s the best dimension size to use?
The default 3,072 dimensions gives peak performance, but the Matryoshka technique means you can reduce to 768 dimensions with minimal quality loss (67.99 vs 68.17 on MTEB). For most production use cases, 768 or 1,536 dimensions offer a strong balance between quality, storage cost, and latency.

Does it replace gemini-embedding-001?
Gemini Embedding 2 is a significant upgrade with multimodal capabilities, but embeddings from the two models are not compatible. Upgrading requires re-embedding your existing data, so it’s not a drop-in replacement.

Top AI Product