Introducing EmbeddingGemma: The Best-in-Class Open Model for On-Device Embeddings

Key Takeaways

[Music] Hi, I'm Alice. >> I'm Lucas. We're product managers at Google DeepMind. >> And today we are incredibly excited to introduce Embedding Gemma, our state-of-the-art embedding model designed for mobile first AI. Embedding Gemma is a 300 million parameter text embedding model designed to power generative AI experiences directly on your hardware. Embeddings are numerical representations of data. This model transforms text like messages, emails or notes into a vector of numbers to represent meaning in a highdimensional space that a generative model can then use for downstream tasks. Embedding Gemma is small, fast, and efficient. Thanks to quantization aware training, you can run the model with as little as 300 megabytes of RAM while preserving state-of-the-art quality. It generates embeddings of 768 dimensions, but thanks to MROSKA representation learning, you can customize the model's output dimensions and go down to 128. Based on the same technology and research that powers our Gemini embedding models, embedding Gemma brings that state-of-the-art capability in a smaller and more lightweight model. Think highquality semantic search, fast and relevant information retrieval, or customized classification and clustering, just to name a few opportunities. Embedding Gemma achieves the best score on the comprehensive massive text embedding benchmark for models under 500 million parameters. The gold standard for text embedding evaluation trained across 100 plus languages. Embedding Gemma brings proven performance to instantly connect with diverse and global audiences. We've engineered embedding Gemma specifically for ondevice performance to ensure efficient computations and minimal memory footprint even on resource constrained hardware. Embedding Gemma facilitates ondevice embedding of local documents. So sensitive user data never leaves the device. And because it works offline, it means Frontier search and retrieval features work regardless of connectivity. Together with our generative models like Gemma 3N, you can build powerful mobile first generative AI experiences and efficient retrieval augmented generation pipelines. This means your applications can now leverage user context from data to provide more personalized and helpful responses such as understanding that you need your carpenters's number for help with damaged floorboards. Here's an example of what embedding Gemma can power. What you are seeing is how a user can utilize embedding Gemma to query previously opened articles or other web pages. The model embeds each page as it's opened in real time. Then with a browser extension that uses embedding Gemma, the user can ask a question to retrieve the contextually relevant articles. And because the embeddings are created on device, all this is happening without leaving the user's hardware. >> And it's designed with customization in mind. fine-tune embedding Gemma for your domain or in a particular language. It works across popular tools and platforms such as hugging face and Kaggle. Check out our notebook examples part of the Gemma cookbook to get started. Our next generation of ondevice embedding models is here and it's open for everyone. It's small, fast, and efficient. Download Embedding Gemma and get started building right now. >> You can find links in the description below. We can't wait to see what Embedding Gemma unlocks for you. [Music]