- Author
- Sahil Dua and Henrique Schechter Vera, Research Engineers, Google DeepMind
- Published
- 6 October 2026
- Kind
- Company announcement post with feature list and links
- Original
- blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2/
- Subject
- Language models and training
- Themes
- Given away
What it is
Two Google DeepMind research engineers, Sahil Dua and Henrique Schechter Vera, wrote the post. It announces EmbeddingGemma 2, a follow-up to a text-only embedding model from last year. The post describes features, uses and where to download it.
What it says
- Google says EmbeddingGemma 2 maps text, images, audio and video into one embedding space.
- The model has 740 million parameters and is released under an Apache 2.0 license.
- It is modular: 270M parameters for text only, with optional vision and audio encoders.
- Matryoshka training lets vectors shrink from 768 to 128 dimensions, saving up to 6x storage.
- The 8K token context covers up to 5.5 minutes of audio or 29 images.
- Google reports MTEB Code rising from 68.76 to 78.68 compared with the first EmbeddingGemma.
- Quantized on a Pixel 11 Pro, it needs about 191MB (text) or 567MB (full).
What evidence it gives
The post gives headline figures for size, memory and context length. It gives one benchmark change with numbers, on code retrieval. It links demo apps for media search. It shows no table for audio or vision scores.
Interests and limits
Google DeepMind built the model and wrote the post. The best-in-class claims are the maker's own and were not independently tested. The linked model card with full results was not opened. Audio and vision scores are described but not shown.