Google releases EmbeddingGemma 2 for local search across text, images and audio

Google launched EmbeddingGemma 2 on October 6, extending its embedding model family beyond text. The model represents code, written content, pictures, video and audio in a common numerical format for search and retrieval.

The complete model contains 740 million parameters, but developers can use a 270-million-parameter configuration for text alone. Optional components add image and audio processing. Google released it with an Apache 2.0 license and an 8,192-token context capacity.

The original EmbeddingGemma focused on text. Its successor supports searches across media formats, including finding video segments with an audio query. Google said pairing it with Gemma 4 enables applications that retrieve local information and generate answers without a cloud connection.

Model weights are available through Hugging Face and Kaggle, with development support from Google's on-device AI tools. Google also plans to add the model to Gemini Enterprise Agent Platform Model Garden, but its announcement gives no availability date.