EmbeddingGemma 2

What it is

EmbeddingGemma 2 builds on Gemma 4 advances: four modalities in one embedding space, plus Matryoshka Representation Learning (MRL)—native 768-d output, truncatable to 512 / 256 / 128-d to cut vector storage cost.

• Native multimodality: text, images, video, audio.
• Multilingual + code: 100+ languages; ~14% improvement on code tasks vs predecessor (per model card).
• Flexible footprint: 270M text backbone + optional vision (170M) and audio (300M) encoders.
• Context: 8K token window; task instruction prefixes steer search / classification / clustering.

Overview from the HF card: 740M total params; native output 768-d; mean pooling; 512→768 projection.

Who made it

Published by Google DeepMind, hosted under the Hugging Face google org as google/embeddinggemma-2. Tagged for feature-extraction with transformers / sentence-transformers / safetensors.

HF cardData lists Apache-2.0; the model card also requires the Gemma Prohibited Use Policy. Always follow the official license and policy text.

What it is for

Intended downstream uses listed on the model card include:

• Retrieval / RAG across text, code, images, video, and audio.
• Classification (sentiment, moderation, audio events, image categories).
• Clustering and near-duplicate discovery across modalities.
• Semantic similarity for recommendations, paraphrase detection, cross-lingual alignment.
• Fact verification via evidence retrieval for a claim.

Text-only effective size ~270M; full multimodal ~740M.

How to try

Official quick start (sentence-transformers):

pip install -U sentence-transformers transformers

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("google/embeddinggemma-2")

query = "What causes the northern lights?"
document = "The northern lights are caused by charged particles from the sun."

query_emb = model.encode(query, prompt_name="SearchQuery")
doc_emb = model.encode(document, prompt_name="Document")
print(model.similarity(query_emb, doc_emb))

Use recommended text task prefixes; do not prefix images/video/audio. For MRL truncation pass truncate_dim and normalize_embeddings=True. Prefer bfloat16 or float32—avoid float16 (NaN risk per model card).

Source of truth: the model card.

vs EmbeddingGemma 1

Selected 768-d full-precision figures from the model card:

• MTEB multilingual v2 Mean(Task): 61.36 vs 61.15 (roughly flat).
• MTEB code v1 NDCG@10: 78.68 vs 68.76 (large code gain).
• New image / video / audio benchmarks (MIEB, MMEB v2, MSEB, MAEB) with no EmbeddingGemma 1 column.

Main deltas: unified four-modality space, optional encoders, and MRL truncation. Numbers belong to the HF card.