Google has released a lightweight multimodal embedding model, including images and audio, that runs entirely on the device.
Google released EmbeddingGemma 2. An embedding model turns text or images into lists of numbers (vectors) so items can be searched by closeness of meaning. This one maps text, code, images, video and audio into a single space. It has 740 million parameters in total and runs text-only at 270 million; adding a 170 million vision encoder and a 300 million audio encoder gives the full version. On a Pixel 11 Pro it needs about 191MB of RAM for text only and about 567MB for the full model. It is Apache 2.0 licensed and available on Hugging Face and Kaggle.
Because it runs on the device, search features can be built without sending data off it. The context window is 8K tokens (the units an AI splits text into), four times the predecessor's, and the MTEB Code score rose 9.92 points to 78.68. These are Google's own figures, and no independent verification was found.
Whether third-party comparisons against similar-sized models appear, and whether the small on-device configurations keep image and audio search quality. With the model only just released, these two points will show how practical it is.