Google has introduced EmbeddingGemma 2, a new open and lightweight AI model designed to bring powerful multimodal search and retrieval capabilities directly to smartphones, laptops and other consumer devices.
Unlike the original EmbeddingGemma, which focused primarily on text, the new model can work with text, code, images, video and audio in a shared embedding space. This means developers can build applications that understand different types of content together without necessarily sending user data to the cloud.
What Is Google EmbeddingGemma 2?
EmbeddingGemma 2 is an open multimodal embedding model based on the Gemma 4 architecture. It has 740 million parameters and is released under the commercially permissive Apache 2.0 license.
The model converts different forms of information into numerical representations called embeddings. These representations allow software to identify relationships between content and perform tasks such as semantic search, retrieval, classification and recommendation.
For example, a developer could build an app that finds a particular video clip using a text description or searches through recorded audio using a voice memo or written query.
Google says the first EmbeddingGemma model has already surpassed 20 million downloads, with developers using it for on-device search and privacy-focused retrieval-augmented generation (RAG) applications.
Multimodal AI in a Smaller Package
One of the main advantages of EmbeddingGemma 2 is its modular architecture.
The text component requires as little as 270 million parameters, while developers can optionally add a 170-million-parameter vision encoder and a 300-million-parameter audio encoder.
This means developers do not necessarily have to load the entire model when their application only needs text or code embeddings.
The model supports text, images, video and audio within the same embedding space, making it possible to build cross-modal search experiences using a single model.
Designed for On-Device AI
Google is positioning EmbeddingGemma 2 specifically for local AI processing.
With quantization, Google says the text-only version can use around 191MB of active RAM, while the complete multimodal version can require around 567MB on a Pixel 11 Pro.
Running AI locally can offer several advantages. Applications can reduce the amount of information sent to remote servers, potentially improving privacy while also reducing network-related delays.
It also means certain search and retrieval functions can continue working without an internet connection.
8K Context Window Is Four Times Larger
EmbeddingGemma 2 also gets a significantly larger context window than its predecessor.
The model supports an 8,192-token context window, which Google says is four times larger than EmbeddingGemma 1.
For multimodal workloads, it can process up to around 5.5 minutes of audio, 29 images or 58 video frames, including combinations of different input types.
This gives developers more flexibility when building applications that need to understand longer pieces of multimedia content locally.
Google Improves Code Search Performance
Google says EmbeddingGemma 2 delivers a substantial improvement in code-related workloads.
Its MTEB Code score rises from 68.76 with the original EmbeddingGemma to 78.68 with the new model, representing a 9.92-point improvement.
This makes the model particularly useful for applications that need to index codebases, perform semantic code searches or help coding agents retrieve relevant sections of a project.
Google also says the model performs strongly across image, video, document and audio benchmarks compared with other sub-1-billion-parameter multimodal embedding models.
Up to 6x Less Vector Storage
Another interesting feature is Matryoshka Representation Learning (MRL).
EmbeddingGemma 2 produces 768-dimensional embeddings by default, but developers can reduce the output to 512, 256 or 128 dimensions depending on their needs.
This flexibility can significantly reduce the amount of storage required for local vector databases. Google says developers can achieve up to a 6x reduction in vector storage and memory requirements.
For applications managing large collections of documents, images, audio or other content, the reduction could make local semantic search considerably more practical.
EmbeddingGemma 2 Can Power Offline Search
One of the most practical uses for the new model is offline semantic search.
Because EmbeddingGemma 2 can generate embeddings locally, applications can search across stored content without constantly sending information to cloud servers.
A user could theoretically search a local media library using a description, locate a specific moment in a video using a text query, or find relevant files based on their meaning rather than their exact filenames.
Google is also showcasing multimodal applications through its AI Edge tools, including media search and video moment discovery.
Works With Gemma 4 for On-Device RAG
EmbeddingGemma 2 can also work alongside Gemma 4 to create local retrieval-augmented generation systems.
In this setup, EmbeddingGemma 2 can retrieve relevant information while Gemma 4 handles the reasoning or generation stage.
Google says both models share technologies including the text tokenizer and audio encoder, allowing developers to build unified pipelines with a lower combined memory footprint.
This could be useful for creating private AI assistants that can search through local documents and other media without requiring all data to be uploaded to a remote service.
Developers Can Download EmbeddingGemma 2 Now
Google has made EmbeddingGemma 2 available through several developer platforms.
The model weights are available on Hugging Face and Kaggle, while versions optimized for on-device deployment are available through LiteRT Community on Hugging Face.
Developers can also use Google AI Edge MediaPipe for embedding and retrieval tasks or LiteRT for custom integrations.
The model can be served using popular tools and frameworks including Transformers, Sentence Transformers, MLX, vLLM, llama.cpp, SGLang, Ollama and LM Studio.
Why EmbeddingGemma 2 Matters
Embedding models are an important part of modern AI applications because they help software understand relationships between pieces of information.
The key difference with EmbeddingGemma 2 is that Google is making these capabilities available in a relatively compact, open model designed for consumer hardware.
That opens the door to more applications where AI-powered search, retrieval and classification can happen directly on phones and laptops rather than relying entirely on cloud infrastructure.
For developers, the combination of multimodal support, flexible model size, offline processing and the Apache 2.0 license could make EmbeddingGemma 2 an attractive option for building privacy-focused AI applications.
Final Verdict
Google EmbeddingGemma 2 is a significant expansion of the company’s lightweight embedding model family.
With support for text, code, images, video and audio, a 740-million-parameter architecture, an 8K context window and efficient on-device inference, the model is designed to bring multimodal AI capabilities closer to everyday consumer hardware.
Its ability to work offline could also make it useful for privacy-sensitive applications, while flexible embedding sizes can help developers reduce storage requirements.
Google’s new model is now available through Hugging Face and Kaggle, giving developers an open option for building the next generation of on-device semantic search, multimodal retrieval and RAG applications.
