Collect Video Testimonials That Build Trust
EmbeddingGemma 2
EmbeddingGemma 2 is Google's sub 1B open multimodal embedding model for semantic retrieval across text, code, images, video and audio in a shared 768 dimensional vector space.
About EmbeddingGemma 2
EmbeddingGemma 2 is a compact open multimodal embedding model based on Gemma 4. It maps text, code, images, video and audio into a shared 768 dimensional embedding space, enabling semantic similarity and retrieval across different content types.
The model uses modular encoders that can be selectively loaded depending on the required modalities. The text and code configuration uses 270M parameters, text plus vision uses 440M parameters, text plus audio uses 570M parameters, and the complete multimodal configuration uses 740M parameters. Its text and code component is based on an adapted Gemma 4 decoder with an 8,192 token context window.
EmbeddingGemma 2 supports cross modal retrieval and interleaved inputs containing combinations of text, images, video and audio. It also supports task specific prompts for text and code embeddings, including separate prompts for search queries and documents.
The model uses Matryoshka Representation Learning to support embedding dimensions of 768, 512, 256 and 128. Google documents support for running the model through sentence-transformers and several additional development and deployment tools. The model weights are released under the Apache 2.0 license.
Key Features
- Multimodal embeddings: Creates embeddings for text, code, images, video and audio in a shared 768 dimensional vector space.
- Cross modal retrieval: Embeddings from different supported modalities can be compared using semantic similarity.
- 8,192 token context: All supported modalities share an 8,192 token context window.
- Modular encoders: Developers can load configurations ranging from the 270M text and code setup to the complete 740M multimodal model.
- Flexible embedding dimensions: Matryoshka Representation Learning supports truncating embeddings from 768 dimensions to 512, 256 or 128 dimensions.
- Interleaved multimodal inputs: A single embedding can represent combinations of text, images, video and audio.
Pricing
Contact the sales team for pricing information.
Use Cases
- Multimodal semantic search across text, images, video and audio
- Retrieval augmented generation document retrieval
- Source code and technical documentation search
- Local and on device content indexing and retrieval
Pros & Cons
Pros:
- Supports text, code, image, video and audio retrieval within one compatible embedding space
- Modular architecture allows unused vision or audio encoders to be excluded from memory
- Embedding dimensions can be reduced to 512, 256 or 128 dimensions for storage constrained applications
Cons:
- All modalities are limited by an 8,192 token shared context window
- Google reports reduced retrieval quality at 128 dimensions, particularly for image, video and speech retrieval
Integrations
sentence-transformers, Hugging Face Transformers, vLLM, SGLang, MLX, Ollama, LM Studio, LiteRT, Unsloth, Google AI Edge Gallery
FAQ
What is EmbeddingGemma 2?
EmbeddingGemma 2 is a sub 1B open multimodal embedding model based on Gemma 4 that maps text, code, images, video and audio into a shared 768 dimensional vector space.
Which modalities does EmbeddingGemma 2 support?
It supports text, code, images, video and audio.
What is the context window of EmbeddingGemma 2?
All supported modalities share an 8,192 token context window.
How large is EmbeddingGemma 2?
Its modular configurations range from 270M parameters for text and code to 740M parameters for the full multimodal configuration. Text plus vision uses 440M parameters, while text plus audio uses 570M parameters.
What embedding dimensions does EmbeddingGemma 2 support?
Its native embedding space is 768 dimensions. Matryoshka Representation Learning allows embeddings to be truncated to 512, 256 or 128 dimensions.
Can EmbeddingGemma 2 perform cross modal search?
Yes. Embeddings generated from supported modalities occupy the same vector space and can be compared using semantic similarity.
Does EmbeddingGemma 2 support mixed multimodal inputs?
Yes. It supports interleaved inputs that can combine text with images, video or audio into a single embedding.
Can developers load only selected modalities?
Yes. Vision and audio encoders can be excluded when loading the model, allowing developers to use smaller configurations for the modalities they need.
What development tools support EmbeddingGemma 2?
Google documents support for sentence-transformers, Hugging Face Transformers, vLLM, SGLang, MLX, Ollama, LM Studio and LiteRT, with Unsloth referenced for efficient fine tuning.
What license is EmbeddingGemma 2 released under?
The model weights are released under the Apache 2.0 license.
Explore product videos and infographics for EmbeddingGemma 2.
View VideosLast edited
October 7, 2026 at 5:25 PM by Venkatraman Chandrasekaran
Similar to EmbeddingGemma 2
