Google releases EmbeddingGemma 2, an open multimodal embedding model
The short version
Google released Apache 2.0-licensed EmbeddingGemma 2, a 740M-parameter model for unified text, code, image, audio, and video embeddings on devices.
What changed
EmbeddingGemma expands from text embeddings to native multimodal embeddings, with weights available on Hugging Face and Kaggle.
What it means for you
Developers can build local semantic retrieval across text, media, and code, but full multimodal deployments must budget for memory and encoder overhead.
Google released EmbeddingGemma 2 on October 6. The Apache 2.0-licensed open model maps text, code, images, audio, and video into one embedding space for privacy-first semantic retrieval on local hardware.
Google states that the model has 740M parameters, with modular text, vision, and audio components. This extends local RAG, media-library search, and multimodal classification to more input types using one embedding model. Teams should measure quality and memory on their target hardware and modality mix. See the Google DeepMind announcement.
Related tools
Fact check
Coverage timeline
- Primary sourceGoogleGoogle
- Primary sourceGoogle DeepMindGoogle DeepMind
- CoverageSimon WillisonSimon Willison