EmbeddingGemma: a 308M multilingual embedding model for devices
Google released EmbeddingGemma, a 308M-parameter embedding model with a 2K context supporting 100+ languages; at release it was the top text-only multilingual model under 500M on MTEB.
What happened
Google released EmbeddingGemma, with integrations including Sentence Transformers, LangChain, LlamaIndex, Haystack, txtai, transformers.js, Text Embeddings Inference and ONNX Runtime, plus a fine-tuning example.
The technical details
From the post:
- 308M parameters and a 2K-token context
- trained for 100+ languages
- at the time of writing, the highest-ranking text-only multilingual embedding model under 500M on the Massive Text Embedding Benchmark (MTEB)
- aimed at on-device use: mobile RAG pipelines and agents
Why it matters
Embeddings are the retrieval half of RAG (my March and May 2026 study posts used the 22M all-MiniLM-L6-v2). A stronger multilingual embedder that still runs on a phone makes private, on-device search more practical.
#embeddings#retrieval#on-device#google
Sources
This is a summary written from the linked public sources. Figures and numbers are as reported there.