# EmbeddingGemma
EmbeddingGemma is Google's small open embedding model: 308M parameters, built on the Gemma 3 backbone, trained on more than 100 languages. It exists to run [[Embeddings|embedding]] workloads on the device instead of in someone else's data centre.
The numbers that matter:
- **308M parameters**, under 200MB of RAM once quantized
- **Sub-22ms** on an EdgeTPU
- **2K token** context window
- **768 dimensions**, truncatable to 512, 256 or 128 thanks to [[Matryoshka Embeddings|Matryoshka Representation Learning]]
- **#1 on MTEB** for multilingual, English and code, among models under 500M parameters. It still places #2 or #3 under 1B, competing with models roughly twice its size.
## Why the Matryoshka part matters
Most embedding models give you one vector size and you live with it. MRL trains the model so the first N dimensions are still a usable embedding on their own. You truncate 768 down to 128 and keep most of the quality.
That turns storage into a dial. A million documents at 768 float32 dimensions is about 3GB; at 128 it's around 500MB. For anything running on a phone or inside a browser, that's the difference between shipping and not shipping.
Truncation is a one-line change in sentence-transformers.
## What it's for
[[Retrieval-Augmented Generation (RAG)]] pipelines, [[Semantic Search]], classification, clustering, mobile-first chatbots. Anything where the corpus is private, the connection is unreliable, or the per-call cost of a hosted embedding API doesn't survive contact with your usage numbers.
Pair it with a small local generator and the whole RAG loop stays on the machine. That's the [[Edge AI]] story in practice, and it's the reason I keep an eye on this size class rather than the headline models.
One caveat: embedding models are prompt-sensitive. EmbeddingGemma expects task prefixes (query versus document, classification, clustering) and skipping them costs you real retrieval quality. Read the docs before benchmarking it against anything.
## Availability
[[HuggingFace]], Kaggle, [[Vertex AI]], sentence-transformers, and the usual local runtimes. Released under the Gemma Terms of Use, which permit commercial use and fine-tuning.
## References
- Documentation: https://ai.google.dev/gemma/docs/embeddinggemma
- Model page: https://deepmind.google/models/gemma/embeddinggemma/
- Announcement: https://developers.googleblog.com/en/introducing-embeddinggemma/
- Technical report: https://arxiv.org/abs/2509.20354
- Hugging Face collection: https://huggingface.co/collections/google/embeddinggemma
- Hugging Face blog post: https://huggingface.co/blog/embeddinggemma
- MTEB leaderboard: https://huggingface.co/spaces/mteb/leaderboard
## Related
- [[Gemma]]
- [[Gemma 4]]
- [[Google]]
- [[Google DeepMind]]
- [[Embeddings]]
- [[Matryoshka Embeddings]]
- [[Semantic Search]]
- [[Retrieval-Augmented Generation (RAG)]]
- [[AI Retrieval Patterns]]
- [[Edge AI]]
- [[Ollama]]
- [[HuggingFace]]
- [[Gemma AI Skills]]