Google DeepMind ships Gemini Embedding 2 — native multimodal RAG across text, image, video, audio, code

Google DeepMind released Gemini Embedding 2 on May 29 via the Gemini API and Google Cloud Vertex AI. Unlike most embedding models, which embed text and then bolt on image or audio encoders, Gemini Embedding 2 is natively multimodal — a single model produces embeddings for text, image, video, audio, documents, and code into a shared vector space. The pitch is straightforward: build a RAG system that searches a video library, a code repository, and a PDF corpus with one index and one query.

Google reports state-of-the-art results across several embedding benchmarks (specific numbers not disclosed in the announcement covered here), positioning the model against OpenAI's text-embedding-3, Cohere Embed v3, and Voyage AI. The release fits a broader Google I/O 2026 agentic pivot: Gemini 3.5 Flash now serves as the default in AI Mode, Personal Intelligence expanded to nearly 200 countries in 98 languages, and Google's analyst pitch (per Ken Huang) frames Gemini 3.5 Flash as a 'distributed agent runtime' across Antigravity, Spark, and Workspace.

Developer reaction is mixed. On Google's own AI dev forums, a widely-shared post titled 'Gemini has become the most unreliable frontier AI — we need fixes not new features' captured frustration with reliability regressions during the rapid release cadence. Meanwhile a Medium analysis arguing Google has five structural advantages over OpenAI — distribution, TPU silicon, data, talent, and integration — circulated widely as developers re-evaluated the leader narrative.

Watch next: independent benchmarks (MTEB-Multimodal, BEIR-image), pricing relative to OpenAI's embeddings, and whether enterprises consolidate their text+image+audio retrieval stacks onto a single Gemini index versus keeping specialist embedding models per modality.