Google released EmbeddingGemma 2 on October 6, extending its embedding model from text to code, images, audio and video. The 740-million-parameter model uses the Gemma 4 architecture and is available under an Apache 2.0 license.
Embeddings represent information in a form that software can compare for similarity. Google describes uses such as finding a video moment from a voice query or locating audio recordings through text, with processing on local hardware.
The model has an 8,000-token context window and optional vision and audio components. Google says a text-only configuration uses 270 million parameters.
Local retrieval, with deployment choices
Google lists weights on Hugging Face and Kaggle, plus support through its AI Edge tools. Its published memory figures depend on quantization and the tested Pixel 11 Pro configuration; they are not universal requirements for every device. Model Garden availability is described as coming soon.
Sources
Google DeepMind announcement, October 6, 2026
Feature image: AI-generated editorial illustration for Guava Insights. Conceptual artwork, not a product photograph or depiction of an actual facility.
