EmbeddingGemma 2 promises multimodal search on a phone

Searching videos with an ordinary text query can work directly on a phone. Google says that EmbeddingGemma 2 translates text, images, videos, and audio into a shared vector space. The focus here is not on generating answers but on finding matches between different types of data.

The model has a shared 768-dimensional space. This makes it possible, for example, to match a text query with a video. The practical value is clear: personal files can be searched locally without sending them to a server.

The size seems suitable for this scenario. According to Google, the text component uses about 191 MB of RAM on a Pixel 11 Pro, while the full multimodal version uses about 567 MB. This is an estimate for a specific device, not a promise of the same speed on every phone.

The limitation in the available description is significant: it contains no search-quality test results or comparisons with other models. It is also unclear how well the model handles different languages and complex queries. Its compactness makes it possible to run the model on a device, but by itself does not indicate how well it will find the required fragment.

Is it more important for you to search personal photos and videos without sending them to the cloud, or to get the most accurate results possible, even if processing takes place on a server? Source: All Articles in a Row / Machine Learning / Habr