Google EmbeddingGemma 2

EmbeddingGemma 2 is an open, lightweight multimodal embedding model from Google DeepMind that natively maps combinations of text, images, audio, and video into a unified embedding space. Described as the most capable model for on-device multimodal embeddings, it lets developers organize, search, and connect information directly on consumer hardware. Rather than sending data to the cloud, it is designed to run locally, so a single query can search across multiple media types at once. It is aimed at developers who want to power smarter on-device search tools and privacy-first retrieval augmented generation (RAG) pipelines, and it expands on the original EmbeddingGemma by moving beyond text alone. The original EmbeddingGemma was introduced to provide a lightweight option for high-quality text embeddings, helping apps organize, search, and connect information directly on consumer hardware. The developer community's response blew past expectations: EmbeddingGemma has surpassed more than 20 million downloads, and builders have used it to power smarter on-device search tools and privacy-first retrieval augmented generation (RAG) pipelines. However, information on modern devices is rarely purely text. EmbeddingGemma 2 responds by expanding beyond text to unify code, images, video, and audio in a shared embedding space. It can help find a specific video clip from a voice memo, or search through hours of audio recordings based on a text query, all processed by a single, natively multimodal model. EmbeddingGemma 2 is best-in-class for its size. It achieves leading scores among sub-1B multimodal embedders across benchmarks such as MTEB (Massive Text Embedding Benchmark) Code and MAEB (Massive Audio Embedding Benchmark), while matching or outperforming many larger models across text, vision, and audio tasks. The model is also modular by design: it requires as little as 270M parameters for text-only workloads, with optional vision (170M) and audio (300M) encoders available for full multimodal support. This modularity means developers can adopt only the capabilities they need rather than paying the cost of the full multimodal model, which is especially valuable when deployment targets have limited resources. Storage efficiency is another core feature. Using Matryoshka Representation Learning (MRL), developers can dynamically truncate output vectors from 768 dimensions down to 512, 256, or 128 dimensions. This provides up to 6x storage reduction for local vector databases and memory usage, which matters greatly on devices where space and memory are constrained. Combined with optimizations for on-device performance, EmbeddingGemma 2 runs efficiently within tight resource constraints. With quantization, on a Google Pixel 11 Pro, it requires as little as ~191MB active RAM for text-only weights and ~567MB for the full multimodal model. EmbeddingGemma 2 is also extended context ready. It features an 8K token context window, which is 4x larger than EmbeddingGemma 1, allowing it to process up to 5.5 minutes of audio, 29 images, 58 video frames, or interleaved combinations thereof directly on local hardware. On quality, EmbeddingGemma 2 matches the strong multilingual text performance of EmbeddingGemma while delivering a significant 9.92-point improvement on code performance in MTEB Code, rising from 68.76 to 78.68, making it well-suited for local codebase indexing, semantic code search, and coding agent retrieval. Across image, video, documents, and audio, it sets a new standard in quality-per-parameter for sub-1B models and even outperforms some specialist models more than twice its size. The distinctive approach behind EmbeddingGemma 2 is that it is built from the same technology as Gemini Embedding models and runs on the Gemma 4 architecture. Released under a commercially permissive Apache 2.0 license, it has 740 million parameters, making it optimal for on-device inference. Because EmbeddingGemma 2 is built on Gemma 4 and shares its text tokenizer and audio encoder, developers can run both models together in a unified pipeline with a lower combined total memory footprint. This design enables on-device RAG pipelines that understand complex multimodal data, letting an embedding model handle retrieval while a generative model handles reasoning within the same resource budget. The benefits for users flow directly from running embeddings locally. Generating embeddings on-device helps ensure data privacy, reduces pipeline latency, and empowers developers to build cross-modal search and retrieval that works entirely offline. Because nothing needs to be sent to a remote service, sensitive data stays on the device, and pipelines avoid network round trips that would otherwise add delay. Developers also gain the ability to pair EmbeddingGemma 2 with generative models such as Gemma 4 to build on-device RAG pipelines that understand complex multimodal data, all within the resource limits of consumer hardware. Concrete use cases are demonstrated in the Google AI Edge Gallery. Instant Media Search lets you use text or an image to find the top matches in your media library based on semantic similarity. Video Moments Finder lets you locate specific moments in video using text or audio queries. In the Google AI Edge Foresight app, EmbeddingGemma 2 handles local file retrieval while Gemma 4 provides contextual reasoning. And the MediaPipe Decision Task API allows developers to create real-time decision engines that leverage multimodal context for classification, routing, and predictive capabilities. EmbeddingGemma 2 is built for developers and builders working on on-device applications. Model weights are available on Hugging Face and Kaggle, with Gemini Enterprise Agent Platform Model Garden availability coming soon, and the LiteRT Community on Hugging Face offers models optimized for on-device use. For deployment, Google AI Edge MediaPipe provides turnkey embedding, retrieval, and decision tasks, while LiteRT supports custom model integration. Developers can build for the browser with transformers.js or WebGPU. The model works with familiar tools including transformers, sentence-transformers, MLX, vLLM, llama.cpp, SGLang, Ollama, and LMStudio, and embedding vectors can be stored with Qdrant. Fine-tuning guidance is available from Unsloth. It is released under the Apache 2.0 license, and resources include a developer guide, documentation, and guides for inference and fine-tuning. In summary, EmbeddingGemma 2 brings best-in-class, natively multimodal embeddings to on-device hardware under an open, commercially permissive license. By unifying text, code, images, video, and audio in a shared space, scaling from as little as 270M parameters up to a 740M-parameter multimodal model, and supporting flexible vector truncation, it enables private, low-latency, offline search and retrieval augmented generation pipelines that developers can run directly on consumer devices.