Google announced EmbeddingGemma 2, an open model for creating on-device embeddings from text, code, images, audio, video, and combinations of them in one shared space. It has 740M total parameters—a 270M text model plus modular 170M vision and 300M audio encoders—and produces 768-dimensional embeddings. (Google DeepMind announcement, Google model card)

Google announced EmbeddingGemma 2, an open model for creating on-device embeddings from text, code, images, audio, video, and combinations of them in one shared space. It has 740M total parameters—a 270M text model plus modular 170M vision and 300M audio encoders—and produces 768-dimensional embeddings. (Google DeepMind announcement, Google model card)

Google says it supports multimodal search, such as finding a video moment with a voice memo, and private, on-device retrieval-augmented generation when paired with Gemma 4. The model is released under Apache 2.0; its weights are available on Hugging Face and Kaggle. (Google DeepMind follow-up, Google announcement)