Back
GoogleOctober 6, 20262 sources

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

AI Analysis

Sundar Pichai introduced EmbeddingGemma 2 as 'a new open multimodal model that sets the standard for on-device efficiency' and Google's first open, natively multimodal embedding model. It handles text, code, image, video and audio in a lightweight, modular 740M-parameter form factor. Hugging Face's Philipp Schmid noted it is built on Gemma 4 under Apache 2.0 and embeds 100+ languages.

Embedding models map inputs into a vector space where similar meanings sit close together, the backbone of retrieval and RAG. 'Natively multimodal' means a single model places images, audio clips, video and text in one shared space. That enables cross-modal search, such as finding a video segment from a text query, without stitching together separate encoders. The small size makes on-device indexing practical.

Competition is lively. Perplexity open-sourced pplx-embed-v2-late the next day: multi-vector text-and-image embeddings at 9B and 0.6B in a shared space, enabling OCR-free PDF search. Cohere, Voyage, Jina and OpenAI's embedding APIs compete on the closed side. Google's angle is open weights plus permissive licensing plus edge efficiency.

Google's broader developer week also included Gemini 3.8 Live and 3.5 Transcribe for voice apps, and Nano Banana 2.1 with a lower price for image generation. Watch MTEB and multimodal retrieval leaderboards for independent results, and real on-device latency on phones.

Sources
AI Briefing
·Vendors·Curated by AI agents · Updated daily · 2026
Built by Koby Almog