EmbeddingGemma 2

Embedding modelOpen weightsCard summarized

The most capable model for on-device multimodal embeddings

Source: EmbeddingGemma 2 model card

Use it via API

Self-host only: no published API price.

Code example Compare

Where to call it

Self-host (weights)google/embeddinggemma-2
On-device—
Google Vertex AI—

Supports

Image input

Full feature support

Good for

Retrieval: document search, RAG, code search, spoken-query search
Classification, clustering and semantic similarity
Fact verification

Overview

Google · 01Embedding

EmbeddingGemma 2

The most capable model for on-device multimodal embeddings

Typical price

—

No published price

Dimensions: 768
Max input: 8,192 tokens shared across modalities
Modes: text embeddings, image embeddings
Released: 6 Oct 2026

EmbeddingGemma 2 is Google DeepMind's open multimodal embedding model, built on the Gemma 4 architecture and released under Apache 2.0. It maps text (including code), images, video and audio into one 768-dimensional space, has 740M parameters in total (270M for text alone) and an 8K-token context, and is designed to run on phones and laptops. Google calls it among the strongest multimodal embedding models under 1B parameters. model card; launch post

google/embeddinggemma-2

Typical price: —
Dimensions: 768 (can be shortened to 512, 256 or 128)
Max input: 8,192 tokens shared across modalities (~29 images, ~58 video frames at 1 fps, or ~327 s of audio)
Languages: 100+ languages (training data in 140+ languages)
Weights: Open · Apache 2.0
Released: 6 Oct 2026

text embeddings
image embeddings
audio embeddings
video embeddings
multimodal embeddings
code retrieval