Reported launch
Google launches EmbeddingGemma 2 for on-device multimodal search

- The model is designed for on-device tasks and is available under an open license.
- EmbeddingGemma 2 has 740 million parameters and can run on devices like the Pixel 11 Pro.
- The model uses modular encoders so developers only load the components they need, with a text-only version using about 191MB of RAM.
- Google claims EmbeddingGemma 2 outperforms competing embedding models up to twice its size on multimodal benchmarks.
Source details
- First seen here
- 6 Oct 2026 16:59 UTC
- Outlets
- 13
- Status
- Google DeepMind announced it · 2026-10-06
