Google DeepMind Releases EmbeddingGemma 2, a 740M Open Multimodal Embedding Model Built on Gemma 4
740M open multimodal embedding model ships today. Five input types, one vector space, Apache 2.0 license.
Why it matters
Google DeepMind releases EmbeddingGemma 2, a production-ready open embedding model that unifies text, image, audio, video, and code into a single 768-dimensional space. Practitioners can now ground retrieval and search across modalities without proprietary APIs.
The key facts
14 to knowModel size: 740M parameters
Output dimension: 768d embedding space
Input types: 5 (text, image, audio, video, code)
License: Apache 2.0 (open-weight)
Built on Gemma 4 base
Release date: October 6, 2026
Shipped today (GA, not preview)
EmbeddingGemma 2: 740M parameters
Unifies 5 input modalities (text, image, audio, video, code)
768-dimensional output space
Built on Gemma 4
Apache 2.0 license (open-weight)
Released October 6, 2026
Google DeepMind source
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: Google DeepMind's EmbeddingGemma 2 maps 5 input types into one 768d space and ships today under Apache 2.0.