Google releases EmbeddingGemma 2 multimodal AI model

Google has released EmbeddingGemma 2, an open-weights multimodal embedding AI model based on the Gemma 4 architecture. The model is designed to integrate text, code, images, video, and audio into a single 768-dimensional vector space, allowing for cross-modal retrieval without the need for prior captioning or transcription.

Released under the Apache 2.0 license, the model features up to 740 million parameters. It is optimized for on-device deployment; testing on a Pixel 11 Pro showed the full multimodal model requires approximately 567MB of active RAM. Developers can use modular encoders to reduce resource usage, such as a 270-million-parameter base for text and code tasks.

Google has also increased the context window from 2,048 to 8,192 tokens. This expansion allows the model to process inputs such as 5.5 minutes of audio or approximately 58 video frames in a single instance. On-device deployment is currently available through LiteRT and MediaPipe Tasks, with Android ML Kit integration expected in the coming weeks.