EmbeddingGemma
2025-09-04 · Publisher : Google · Domain : Artificial Intelligence
Google releases EmbeddingGemma, an open embedding model with 308 million parameters for on-device use: first on the MTEB ranking among multilingual models under 500M.
What changes
- Best open multilingual embedding model under 500 million parameters on the MTEB ranking, according to Google.
- Trained on more than 100 languages, based on the Gemma 3 architecture.
- Customizable outputs from 768 to 128 dimensions thanks to Matryoshka Representation Learning.
- Under 15 ms to produce an embedding (256 tokens) on EdgeTPU.
- QAT quantization: under 200 MB of memory; same tokenizer as Gemma 3n.
### Specifications - Parameters: 308 M (about 100 M for the model and 200 M for the embeddings). - Context: 2,000 tokens (2K). - Use cases: RAG and offline semantic search, with Gemma 3n. - Availability: open weights on the usual platforms. ### Pricing Open model, with no usage pricing of its own. ### Benchmarks MTEB ranking: best open multilingual model under 500 M (rank published by the vendor, score not included in the text). ### Key takeaways For building private semantic search on a computer or mobile device, without sending documents to the cloud.
Links
Tags : Open Source, LLM, Google, Gemma, Embeddings