EmbeddingGemma 2 puts several media types in one retrieval space

Google’s EmbeddingGemma 2 produces comparable vectors for text, code, images, audio and video, enabling search across different media types. Its modular design lets a text-only application omit the vision and audio components. The released weights use Apache 2.0, with integrations intended to support local and on-device retrieval as well as larger deployments.

Google DeepMind · Oct 6, 2026

Why it matters

Teams building private search can evaluate one embedding space instead of maintaining disconnected indexes for each input type.

More context

Try the available Hugging Face or Kaggle weights against your own retrieval examples. Vector sizes can be reduced from 768 dimensions to save storage, with quality measured at each setting. This model supplies retrieval representations; it does not itself generate answers.

Keep in mind

Google’s benchmark and device-memory figures depend on its evaluation and configuration. We have not tested them. Multimodal components add resource requirements, and announced future distribution options should not be treated as available today.

Original read