Gemini Embedding 2: Google’s First Native Multimodal Embedding Model

What Embedding Models Do

Embedding models convert any content—text, image, video, audio, or document—into a numeric vector so that semantic similarity can be measured by distance in vector space; similar items are close, unrelated items are far apart.

Five Modalities in One Vector Space

Gemini Embedding 2 supports:

Text : up to 8,192 tokens, 100+ languages.

Image : up to 6 PNG or JPEG images per request.

Video : up to 120 seconds (80 seconds if audio is included), MP4 or MOV.

Audio : up to 80 seconds, MP3 or WAV, processed natively without transcription.

Document : PDF files up to 6 pages, OCR supported.

The modalities can be interleaved; for example, a text string together with two images is processed as a single, unified input and yields one combined semantic vector.

Core Technology: Matryoshka Representation Learning

Gemini Embedding 2 introduces Matryoshka Representation Learning (MRL), a “nested‑vector” approach where a large vector contains progressively smaller vectors of decreasing precision. Users can select the output dimension that matches their performance and storage needs:

3,072 dim (default) – highest accuracy for recall‑critical scenarios.

1,536 dim – a balance point; benchmarks show it can be slightly better than 2,048 dim.

768 dim – lightweight, lowest storage cost, suitable for large‑scale deployment.

The model also accepts custom task instructions, e.g., task:code retrieval, to steer the embedding toward specific objectives.

Performance Highlights

According to Google, Gemini Embedding 2 sets new industry standards across several dimensions:

Text : ranks in the top 5 of the MTEB multilingual leaderboard.

Image & Video : achieves state‑of‑the‑art results among commercial closed‑source models.

Audio : introduces strong speech embeddings, a capability absent in prior mainstream models.

Multilingual : supports 100+ languages with cross‑language semantic alignment that outperforms previous Gemini versions.

New Scenarios Enabled by a Unified Vector Space

The unified space unlocks applications that were previously blocked by modality boundaries:

Multimodal Semantic Search : upload an image and retrieve related videos or articles.

Multimodal RAG : knowledge bases containing text, charts, and audio can be queried uniformly.

Cross‑Language Document Intelligence : 100+‑language PDFs are vectorized for multilingual retrieval.

Content Recommendation : recommend articles or podcasts based on video semantics.

Sentiment Analysis & Clustering : unified analysis of multilingual, multimedia content.

Voice‑Only Retrieval : audio is embedded directly without transcription.

Integration with Vector Databases and Orchestration Frameworks

The API returns a single vector that can be stored in Qdrant, Weaviate, ChromaDB, or Google Vector Search. Deep integrations are provided for LangChain, LlamaIndex, and Haystack.

Three‑Line Code to Get Started

result = client.models.embed_content(
    model="gemini-embedding-2-preview",
    contents=[
        "What is the meaning of life?",  # text
        image_bytes,                     # image
        audio_bytes                      # audio
    ]
)

Availability

Gemini Embedding 2 is in Public Preview. Model ID: gemini-embedding-2-preview. It is reachable via the Gemini API and Vertex AI, currently in the us‑central1 region (other regions will follow). Billing is standard pay‑as‑you‑go; provisioned throughput and batch prediction are not yet supported. Knowledge cutoff: November 2025.

Conclusion

Embedding models are the foundation of AI applications, but they have long been text‑centric. Gemini Embedding 2 upgrades this foundation to a “full‑modal” base, giving semantic search, RAG, recommendation, and other downstream systems native understanding of images, video, and audio without any extra conversion steps.