On March 10, 2026, Google DeepMind pushed Gemini Embedding 2 into Public Preview: the company’s first natively multimodal embedding model, built on the Gemini architecture. Text, images, video, audio, and PDF documents all map into a single, unified 3072-dimensional vector space covering more than 100 languages. The announcement was signed by product manager Min Choi and Distinguished Engineer Tom Duerig.
For developers, this patches the most fragmented layer of retrieval engineering. Multimodal search used to mean stitching one embedder for text, another for images, and a transcription hop for audio. Google’s framing is blunt: one space replaces a whole assembly of single-modality models, cutting complexity and embedding cost at the same time.
One Model, One Space, Five Modalities
The input spec decides what actually fits in the space:
- Text input up to 8,192 tokens; up to 6 images per request (PNG, JPEG).
- Video up to 120 seconds (MP4, MOV); PDFs up to 6 pages.
- Audio is ingested natively, with no intermediate text transcription — historically the most awkward stage of any voice-retrieval pipeline.
- Interleaved input is supported: a single request can mix images and text, and the model understands the combination natively.
- Google claims performance ahead of leading models on text, image, and video tasks.
Those limits reveal the positioning. This model handles retrieval units — a document page, a short clip, one image — not unbounded multimedia files. The 8,192-token text ceiling sits far below the million-token contexts of the main Gemini models: embedding and generation each do their own job, without pretending to be each other.
3072 Dimensions and Matryoshka Truncation
The default output is 3072 dimensions, built on Matryoshka Representation Learning: information is “nested” inside the dimensions, so vectors can be truncated dynamically without retraining. Google recommends three settings: 3072, 1536, or 768 dimensions.
This is the most practical cost lever in the whole design. Storage, memory, and search cost in a vector database scale roughly linearly with dimensionality: a 1536-dim index over tens of millions of records halves the footprint, and 768 halves it again. Matryoshka turns the quality-versus-cost tradeoff from a one-time engineering decision into a runtime parameter — evaluate retrieval quality offline at full width, then truncate to your latency budget in production, from the same source index.
The API and Ecosystem, Ready on Day One
- The model ID is gemini-embedding-2-preview, live on both the Gemini API and Vertex AI.
- LangChain, LlamaIndex, Haystack, Weaviate, Qdrant, ChromaDB, and Vertex AI Vector Search are all in the first wave of integrations, with Colab notebooks alongside.
Be clear-eyed about where the switching cost actually sits. Replacing an embedding model was never about the API call — it is the index rebuild: change the embedding space and you recompute everything. What day-one framework support lowers is the threshold to try it. Point your existing embedding function at the new model ID, run your eval suite, and you will know whether a full reindex is justified before committing to one.
What It Changes for RAG and Agent Pipelines
Three consequences. First, cross-modal retrieval becomes a native operation: query a video segment with a sentence, or find the original document from a screenshot, without maintaining per-modality indexes and alignment logic. Second, the agentic multimodal RAG pattern Google demos in its developers blog is the most direct application — an agent that reads screenshots, documents, and voice memos at once, with memory unified in one vector representation, so both tool selection and recall sit on a single space. Third, watch the lock-in effect: if the unified space holds up in quality, the gravitational pull of keeping your data and indexes inside Google’s ecosystem grows. Embedding spaces cannot be mixed, so switching models means a full reindex — price that into any migration decision.
The same day, Google also shipped Gemini deeper into Workspace — drafting from your profile, spreadsheet autofill, and Drive Q&A. Infrastructure components and end-user applications moving in the same week sends its own signal: when even “plumbing-grade” pieces like embeddings join the regular release cadence, multimodal retrieval is being treated as part of the Gemini platform’s foundation, not an optional extra.
Sources
- Gemini Embedding 2: Our first natively multimodal embedding model — Google
- Building with Gemini Embedding 2: Agentic multimodal RAG — Google Developers Blog
- Gemini Embedding 2 preview model — Gemini API documentation
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.