EmbeddingGemma 2
Available nowCompact multilingual text retrieval with Google’s open EmbeddingGemma 2 model.
What it does
EmbeddingGemma 2 turns queries and document passages into comparable vectors. This service exposes the text retrieval tower of Google’s open model, hosted by DeepInfra through AIRuntime.
Choose 128, 256, 512 or 768 dimensions. Query and document prefixes are applied by the service, and every output is L2-normalized.
Language coverage
Google documents multilingual retrieval across more than 100 languages. Evaluate retrieval quality on your own collection.
Official language coverage ↗Where it helps
Multilingual knowledge
Find relevant text across a multilingual collection.
Compact indexes
Evaluate smaller vectors when storage matters.
Open model evaluation
Compare Google’s open retrieval model on representative queries.
When to choose it
Use it for text retrieval when you want a compact open model. Evaluate quality on your own languages and documents.
This entry point accepts text only. Each request includes up to 8 unique texts, at most 8,000 UTF-8 bytes each. Do not automatically retry an uncertain request. Cached vectors are free.
Official sources
Paraphrased from the provider’s documentation and original model information. Selection advice is our editorial guidance.
Google · EmbeddingGemma 2 model card ↗DeepInfra · API and pricing ↗Reviewed October 23, 2026