Google's EmbeddingGemma 2 brings on-device search across text, images, audio and video
3-line summary
- Google DeepMind released EmbeddingGemma 2, a lightweight embedding model that maps text, images, audio and video into one space.
- At 740 million parameters it runs on phones, and it uses the commercially permissive Apache 2.0 license.
- It is available now on Hugging Face and Kaggle, and can power fully offline search, such as finding a video clip from a voice memo.
Google DeepMind released EmbeddingGemma 2 on October 6. Embeddings turn the meaning of text or images into lists of numbers so similar things can be found, and the new model puts not just text and code but images, video and audio into a single space. The first version, released last year, has passed 20 million downloads.
Built on the Gemma 4 architecture, it has 740 million parameters, but because the vision and audio encoders are optional, text-only use needs as little as 270 million. Its 8K-token context is four times larger than before, enough for 5.5 minutes of audio, 29 images or 58 video frames.
With quantization, Google says it runs on a Pixel 11 Pro in about 191 MB of memory for text only and about 567 MB for the full multimodal model. Output vectors can be shortened from 768 to as few as 128 dimensions, cutting storage by up to six times. Its code-search score on MTEB Code rose from 68.76 to 78.68.
The claim that it leads models under 1 billion parameters rests on benchmarks Google selected. The model is on Hugging Face and Kaggle and runs in tools such as transformers, Ollama and llama.cpp.
Why it matters
Embeddings, the foundation of search and RAG, now cover photos, video and audio and are small enough to run on a phone. That should make it easier to build apps that search private data, such as photo libraries or voice recordings, without it leaving the device.
Who it's for
- Developers
- If you're building on-device search or RAG, start with the 270-million-parameter text-only setup, and shorten vectors to trade storage against accuracy.
- Beginners
- If you've wondered what embeddings do, "find a video clip from a voice memo" is a good picture of it. The glossary entry on embeddings explains more.
Sources
- EmbeddingGemma 2: an open, lightweight multimodal embedding model — Google DeepMind
