Gemini Embedding 2

Google released Gemini Embedding 2 as its first natively multimodal embedding model, mapping text, images, video, audio, and documents into a unified embedding space for over 100 languages. It launched in public preview on March 10, 2026, via the Gemini API and Vertex AI, expanding on a prior text-only foundation with Gemini-architecture multimodal understanding. The model supports up to 8,192 input tokens, six images per request, 120 seconds of MP4 or MOV video, natively ingested audio without intermediate transcription, and PDFs up to six pages. It handles interleaved inputs and is intended for RAG, semantic search, sentiment analysis, and clustering. Authored by Google's Gemini Embedding Team and DeepMind staff.

Official Gemini API documentation names gemini-embedding-2 as the latest and first multimodal embedding model in the Gemini API, mapping text, images, video, audio, and documents into a unified embedding space for cross-modal search and clustering across 100-plus languages. The text-only gemini-embedding-001 remains available for non-multimodal workloads. The docs cover task-type prefixes to tune retrieval quality, Matryoshka Representation Learning for adjusting output dimensionality, and guidance for migrating from gemini-embedding-001. Sections describe embedding images, audio, video, PDFs, aggregation strategies, and batch embeddings for higher-throughput pipelines.

The HTML version of the same arXiv paper confirms the Gemini Embedding Team authorship and the unified multimodal embedding approach for video, audio, image, and text. It details the multi-task multi-stage contrastive training regime that drives the model's cross-modal generalization on interleaved inputs. Benchmark numbers match the abstract: 62.9 R@1 on MSCOCO, 68.8 NDCG@10 on Vatex, 69.9 on MTEB multilingual, and 84.0 on MTEB Code, all described as state-of-the-art at submission. Tom Duerig and Mojtaba Seyedhosseini are listed among the senior contributors alongside Min Choi and the broader Google DeepMind team.

The arXiv paper 'Gemini Embedding 2: A Native Multimodal Embedding Model from Gemini,' submitted May 26, 2026, introduces the model as a unified-embedding architecture for video, audio, image, and text from the Google Gemini Embedding Team. It applies large-scale contrastive learning in a multi-task multi-stage training pipeline to generalize across interleaved inputs. Reported benchmark scores are 62.9 R@1 on MSCOCO, 68.8 NDCG@10 on Vatex, 69.9 on MTEB multilingual, and 84.0 on MTEB Code, surpassing specialized models on unimodal, cross-modal, and multimodal retrieval. The paper is authored by 87 contributors led by Madhuri Shanbhogue, Zhe Li, and Shanfeng Zhang.

Google Cloud documents Gemini Embedding 2 as a multimodal embedding generator producing 3072-dimensional float vectors by default, with Matryoshka Representation Learning support for smaller output sizes. Inputs cover images, text, documents, audio, and video, mapped into a unified semantic space for cross-modal retrieval such as text-to-image search. The model adds custom task instructions, document OCR, and audio track extraction from video inputs. Maximum input tokens are 8,192, batch inference is supported, and standard pay-as-you-go pricing is available while provisioned throughput is not. The model identifier is gemini-embedding-2.

Google announced the General Availability of Gemini Embedding 2 via the Gemini API and Gemini Enterprise Agent Platform around April 30, 2026, transitioning from the March public preview. The model unifies text, images, video, audio, and PDF documents in a single semantic space across 100-plus languages. Single-call input limits are 8,192 text tokens, 6 images, 120 seconds of video, 180 seconds of audio, and 6 PDF pages, with task prefixes such as task:code retrieval or task:search result to optimize embeddings for agentic multimodal RAG. Separate per-item embeddings are available through the Batch API for query or document indexing.