How to load EmbeddingGemma 2

Per the Google model card (2026-10-06): `pip install -U sentence-transformers transformers`, then `SentenceTransformer("google/embeddinggemma-2")`. For search, encode the query with `prompt_name="SearchQuery"` and documents with `prompt_name="Document"` (Document applies `title: none`; titled docs should be formatted as `title: {title} | text: {content}`). Prefer `torch.bfloat16` when supported, else `float32`—never float16.

Related search phrases like “embeddinggemma-300m” usually point at the earlier text EmbeddingGemma release on Hugging Face, not EmbeddingGemma 2. Prefer the `embeddinggemma-2` repo for multimodal weights. Community GGUF / quantized packs may appear under third-party namespaces—verify license and hash against Google’s card before production. This page does not redistribute weights.

Limitations

HF download size and mirror availability vary by region. Kaggle and Gemini Enterprise Model Garden availability are described as coming soon on Google’s announcement—recheck vendor posts. We are not Hugging Face or Google support.

Need multimodal details?

Token budgets, interleaved inputs, selective encoders.