Today, Google announced open source a newOpen embedding model EmbeddingGemma. This model is small and broad, with308 million parameters, designed forDevice-side AIDesigned to support the deployment of retrieval enhanced generation (RAG), semantic search and other applications on laptops, mobile phones and other devices.
A major feature of EmbeddingGemma is its ability toGenerate high-quality embedding vectors with good privacy, even inDisconnectedIt can run normally even under normal circumstances, and its performance is on par with Qwen-Embedding-0.6B, which has twice the size.
▲Screenshot of Hugging Face open source page
Hugging Face address:
https://huggingface.co/collections/google/embeddinggemma-68b9ae3a72a82f0562a80dc4
According to Google, EmbeddingGemma has the following highlights:
1. Best in class:On the Massive Text Embedding Benchmark (MTEB), EmbeddingGemma ranks highest among open multi-language text embedding models under 500M. EmbeddingGemma is built on the Gemma 3 architecture and has been trained for more than 100 languages. It is small and can run on less than 200MB of memory after quantification.
▲MTEB score: EmbeddingGemma performance is comparable to top models twice the size
2. Designed for flexible offline work:Small, fast, and efficient, it provides customizable output size and a 2K token context window, and can run on everyday devices such as mobile phones, laptops, and desktops. It is designed to work with Gemma 3n, working together to unlock new use cases for mobile RAG pipelines, semantic search, and more.
3. Integrate with popular tools:To make it easy for users to get started with EmbeddingGemma, it already works with users’ favorite tools such as sentence-transformers, llama.cpp, MLX, Ollama, LiteRT, transformers.js, LMStudio, Weaviate, Cloudflare, LlamaIndex, LangChain, and more.
1. It can generate high-quality embedding vectors, and the end-side RAG generates more accurate answers.
EmbeddingGemma will generate embedding vectors. In the context of this article, it can convert text into numerical vectors and represent text semantics in high-dimensional space; the higher the quality of the embedding vector, the better the representation of language nuances and complex features.
▲EmbeddingGemma will generate embedding vectors
There are two key stages in building a RAG process: retrieving relevant context based on user input, and generating well-founded answers based on that context.
To implement the retrieval function, users can first generate an embedding vector of the prompt word, and then calculate the similarity between this vector and the embedding vectors of all documents in the system - in this way, the text fragments most relevant to the user's query can be obtained.
Users can then feed these text snippets, along with the original query, into a generative model such as Gemma 3 to generate contextually relevant answers. For example, the model can understand the phone number you need to call a carpenter to fix a damaged floor.
For this RAG process to be effective, the quality of the initial search step is critical. Poor quality embedding vectors can result in irrelevant documents being retrieved, resulting in inaccurate or meaningless answers.
The performance advantage of EmbeddingGemma is reflected in this - it can provide high-quality (text) representation and provide core support for accurate and reliable end-side applications.
2. Qwen-Embedding-0.6B, which has a performance that is nearly doubled in size but is small in size.
EmbeddingGemma provides state-of-the-art text understanding capabilities commensurate with its scale, with particularly strong performance in multilingual embedding generation.
Compared with other popular embedding models, EmbeddingGemma performs well on tasks such as retrieval, classification, and clustering.
EmbeddingGemma has completely caught up with the gte-multilingual-base model of the same size in the Mean (Task), Retrieval, Classification, Clustering and other tests. Its test results are also close to Qwen-Embedding-0.6B, which is twice the size.
▲Evaluation status of EmbeddingGemma
The EmbeddingGemma model has 308M parameters, mainly composed of approximately 100M model parameters and 200M embedding parameters.
To achieve greater flexibility, EmbeddingGemma leverages Matryoshka Representation Learning (MRL) to provide multiple embedding sizes in a single model. Developers can use the full 768-dimensional vector for best quality, or truncate it to smaller dimensions (128, 256, or 512) to increase speed and reduce storage costs.
Google has shortened the embedded inference time (256 input tokens) to <15ms on EdgeTPU, breaking the speed boundary, which means that the user's AI function can provide real-time response and achieve smooth and instant interaction.
Using quantization-aware training (QAT), Google significantly reduced RAM usage to under 200MB while maintaining model quality.
3. It can be used when disconnected from the Internet and can run on less than 200MB of memory.
EmbeddingGemma enables developers to build flexible and privacy-focused on-device applications. It generates document embeds directly on the device hardware, helping to keep sensitive user data secure.
It uses the same tokenizer as Gemma 3n for text processing, thus reducing the memory footprint of RAG applications. Users can use EmbeddingGemma to unlock new features, such as:
Search personal files, texts, emails, and notifications all at once without an Internet connection.
Personalized, industry-specific and offline support chatbots with RAG and Gemma 3n.
Classify user queries into relevant function calls to aid mobile agent understanding (user needs).
The image below is an interactive demonstration of EmbeddingGemma, which visualizes text embedding in three-dimensional space. The model runs entirely on the device.
▲Interactive demonstration of EmbeddingGemma (Source: Joshua from the Hugging Face team)
Demo experience address:
https://huggingface.co/spaces/webml-community/semantic-galaxy)
Conclusion: Small size and great capabilities accelerate the development of end-side intelligence
The launch of EmbeddingGemma marks a new breakthrough for Google in miniaturization, multi-language and device-side AI. Not only is it close to larger models in performance, it also balances speed, memory, and privacy.
In the future, as applications such as RAG and semantic search continue to spread to personal devices, EmbeddingGemma may become an important cornerstone in promoting the popularization of device-side intelligence.