Summary: Gemini Embedding 2 entered Public Preview on March 10, 2026 as Google’s first natively multimodal embedding model, placing text, images, video, audio, and PDFs in one 3072-dimensional space. Matryoshka truncation supports 3072, 1536, and 768 dimensions, but changing models still requires rebuilding the index. Google claims better performance on text, image, and video tasks.
Gemini Embedding 2 entered Public Preview on March 10, 2026, in an announcement from Google DeepMind. Built on the Gemini architecture, it is the company’s first embedding model that is multimodal by design. Text, images, video, audio, and PDF documents now share one 3072-dimensional vector space that spans over 100 languages. Product manager Min Choi and Distinguished Engineer Tom Duerig put their names on the announcement.
Retrieval engineering has long had a ragged edge, and this is where it gets patched. Multimodal search once meant picking one model for text, another for images, and pushing audio through a transcription step first. A single model can now stand in for that whole assembly, and complexity and embedding cost fall together.
Five Modalities in a Single Space
What can actually go into the space is set by the input spec:
- Text runs up to 8,192 tokens; each request takes at most 6 images, as PNG or JPEG.
- Video is capped at 120 seconds in MP4 or MOV; a PDF may span at most 6 pages.
- Audio goes in natively, so the intermediate text transcription — long the clunkiest stage of a voice-retrieval pipeline — simply disappears.
- Interleaved input works: images and text can share one request, with the combination understood natively by the model.
- Google states that on text, image, and video tasks it outperforms the leading models available today.
Read together, the limits define the job. This model is built for retrieval units — one document page, one short clip, a single image — rather than multimedia files of unbounded size. Its 8,192-token text ceiling also sits nowhere near the million-token contexts of Gemini’s main models: embedding and generation stay in their own lanes.
Matryoshka Truncation on a 3072-Dimension Base
Out of the box the model emits 3072 dimensions, with Matryoshka Representation Learning underneath: information sits “nested” across the dimensions, so a vector can be cut down on demand without any retraining. Three sizes earn Google’s recommendation — 3072, 1536, and 768 dimensions.
Dimensionality is the most usable cost dial in this design. In a vector database, storage, memory, and query cost all grow roughly in line with the dimension count: moving a tens-of-millions-record index to 1536 dimensions cuts its footprint in half, and 768 cuts it in half once more. With Matryoshka, the tradeoff between quality and cost stops being a one-time engineering decision and becomes a parameter — measure retrieval quality offline at full width, then decide in production how far to truncate based on your latency budget.
Ecosystem and API Integration from Day One
- It ships under the model ID gemini-embedding-2-preview, available through the Gemini API and Vertex AI alike.
- The first wave of integrations covers LangChain, LlamaIndex, Haystack, Weaviate, Qdrant, ChromaDB, and Vertex AI Vector Search, with Colab notebooks alongside.
Know where the real switching cost lives. Swapping an embedding model has never been hard at the API-call level — the burden is the index rebuild, because once the embedding space changes, every record has to be recomputed. Day-one framework support lowers something else: the barrier to experimenting. Aim your current embedding function at the new model ID, run your evals, and the results tell you whether the full reindex is worth it before you start one.
The Practical Effect on RAG and Agent Pipelines
Three effects stand out. Cross-modal retrieval turns into a native operation first: a sentence retrieves a video segment, a screenshot leads back to the source document, and no per-modality index or alignment logic needs maintaining. Next, the agentic multimodal RAG pattern shown on Google’s developers blog is the application that maps most directly — an agent taking in screenshots, documents, and voice memos together, its memory unified in one vector representation, so both tool selection and recall run on one space. Finally, keep an eye on lock-in: should the unified space prove itself on quality, the pull toward leaving your data and indexes inside Google’s ecosystem only strengthens. Embedding spaces cannot be combined, so any model change forces a full reindex — make sure migration math includes that cost.
On that same date, Google pushed Gemini further into Workspace — drafting from your profile, spreadsheet autofill, and Drive Q&A. Having infrastructure components and end-user features land in one week carries a message of its own: once “plumbing-grade” pieces such as embeddings ride the routine release train, multimodal retrieval counts as part of the Gemini platform’s foundation rather than an add-on.