Google DeepMind releases EmbeddingGemma 2, its first open natively multimodal embedding model

EmbeddingGemma 2 extends Google's open embedding line from text to full multimodality. According to Sundar Pichai's announcement post (5,400+ likes), it is Google's first open, natively multimodal embedding model. It handles text, code, image, video and audio tasks in a lightweight, modular 740-million-parameter form factor designed for on-device efficiency.

Embedding models power retrieval, semantic search, deduplication and RAG pipelines. A single model that maps several modalities into a shared vector space lets developers search video by text, match images to documents, or retrieve audio clips without stitching together separate encoders. At 740M parameters it can run on phones and laptops, keeping private data on-device for retrieval. The 'modular' framing suggests developers can load only the components they need.

The competition includes proprietary multimodal embedding APIs from OpenAI, Cohere and Voyage, plus open text-focused models from the Qwen and BGE families. Few open multimodal embedders are this small, which gives Google a strong position in on-device RAG, an area that complements its Gemini Nano edge strategy.

The Hacker News reaction (217 points) was positive but practical. Developers want independent retrieval benchmarks across modalities and quantized builds. Watch for MTEB-style results, integration into frameworks such as LlamaIndex and LangChain, and Android system-level adoption.