Google DeepMind's open multimodal embedding model mapping text, images, video and audio into one 768-dim space for on-device use.
Hugging Face model page
1 / 3Overview
What is EmbeddingGemma 2?
An open, lightweight multimodal embedding model.
EmbeddingGemma 2 is Google DeepMind's open multimodal embedding model that maps text (including code), images, video and audio into one shared 768-dimensional embedding space, so developers can run cross-modal retrieval in a single model. At about 740M total parameters it is built for on-device use: each encoder can be loaded selectively (the default full set is around 300M), and it runs on consumer hardware like phones and laptops with low latency. Matryoshka (MRL) training supports 128/256/512/768-dimensional embeddings, cutting vector storage up to 6x. An 8K-token context window covers long documents for RAG, task-specific instruction prefixes steer representations toward search, classification, clustering or similarity, and it understands 100+ languages with stronger code-task performance than its predecessor. It is fully open under Apache 2.0 — free to download, deploy and use commercially — and works with sentence-transformers, Hugging Face Transformers, Transformers.js and the HF Inference API. Made for developers and app builders who need lightweight, on-device multimodal embeddings for search, retrieval-augmented generation, recommendations, classification and clustering.
Platforms and languages
EmbeddingGemma 2 Availability
Platforms
Languages
Capabilities
EmbeddingGemma 2 Key Features
Native multimodality
Map text (including code), images, video and audio into one shared 768-dimensional embedding space for cross-modal retrieval.
Multilinguality and code
Understand 100+ languages and perform code tasks about 14% better than its predecessor.
Flexible on-device footprint
740M total parameters with selectively loadable encoders; the default full set is about 300M, designed for phones, laptops and other consumer hardware.
Elastic embedding dimensions
Matryoshka (MRL) support for 128/256/512/768-dimensional embeddings, reducing vector storage up to 6x with minimal quality loss.
8K context with task steering
8K-token context window for long-document retrieval and RAG, with task-specific instruction prefixes that steer embeddings toward search, classification, clustering or similarity.
Best for
Who uses EmbeddingGemma 2?
Developers and app builders needing lightweight multimodal embeddings
Build search, RAG, recommendation, classification and clustering features that run on-device
Plans and access
EmbeddingGemma 2 Pricing
Free
Free: open weights under Apache 2.0, download and use commercially at no cost.
Common questions
EmbeddingGemma 2 FAQs
Is it free and can I use it commercially?
Yes. It is released under Apache 2.0, so you can download the weights and use them commercially at no cost.
Can I run it on-device or locally?
Yes. It is designed for consumer hardware: you can load only the encoders you need, and the default full configuration is about 300M parameters, suitable for phones and laptops.
Which modalities does it support?
Text (including code), images, video and audio — all mapped into one shared 768-dimensional embedding space for cross-modal search.
How do I use the model?
Through sentence-transformers, Hugging Face Transformers, Transformers.js, or the HF Inference API.
Reviews
See what the community thinks and share your experience.
—
Based on 0 ratings
Rating distribution
Community reviews
No reviews
No written reviews yet.
