Google EmbeddingGemma 2

EmbeddingGemma 2 is an open, lightweight multimodal embedding model from Google DeepMind. It natively maps combinations of text, code, images, audio, and video into a unified embedding space, so a single query can search across all of them. The model is intended for developers building on-device applications: apps that organize, search, and connect information directly on consumer hardware. Released under the commercially permissive Apache 2.0 license, EmbeddingGemma 2 has 740 million parameters, making it optimal for on-device inference, and its model weights are available on Hugging Face and Kaggle. The original EmbeddingGemma was introduced to give developers a lightweight option for high-quality text embeddings. That release was used to power smarter on-device search tools and privacy-first retrieval augmented generation (RAG) pipelines, and it surpassed 20 million downloads as the developer community's response blew past expectations. EmbeddingGemma 2 addresses the next step: content is no longer only text. Users have voice memos, image libraries, video archives, and code repositories, and cloud embedding services can add latency and raise privacy concerns because data has to leave the device. EmbeddingGemma 2 unifies those modalities in a shared space so developers can build cross-modal search and retrieval that works entirely offline. EmbeddingGemma 2 is described as the most capable model for on-device multimodal embeddings. It is built from the same technology as Gemini Embedding models and is positioned as best-in-class for its size, achieving leading scores among sub-1B multimodal embedders on benchmarks such as MTEB (Massive Text Embedding Benchmark) Code and MAEB (Massive Audio Embedding Benchmark). According to Google, it matches or outperforms many larger models across text, vision, and audio tasks, and across image, video, documents, and audio it sets a new standard in quality-per-parameter for sub-1B models, even outperforming some specialist models more than twice its size. It also matches the strong multilingual text performance of the original EmbeddingGemma while improving code performance significantly: the MTEB Code score rises from 68.76 to 78.68, a 9.92-point improvement. That makes the model well suited to local codebase indexing, semantic code search, and coding agent retrieval. Modular by design, EmbeddingGemma 2 requires as little as 270M parameters for text-only workloads, with optional vision (170M) and audio (300M) encoders available for full multimodal support. Developers can therefore scale the deployed footprint to the workload instead of always shipping the largest configuration. Storage efficiency is handled through Matryoshka Representation Learning (MRL): output vectors can be dynamically truncated from 768 dimensions down to 512, 256, or 128 dimensions, which provides up to 6x storage reduction for local vector databases and reduces memory usage. For apps that keep embeddings on the device, this directly affects how much media and content can be indexed and searched locally. The model is optimized for on-device performance and runs efficiently within tight resource constraints. With quantization, on a Google Pixel 11 Pro, EmbeddingGemma 2 requires as little as approximately 191MB of active RAM for text-only weights and roughly 567MB for the full multimodal model. It also features an 8K token context window, four times larger than EmbeddingGemma 1, which allows it to process up to 5.5 minutes of audio, 29 images, 58 video frames, or interleaved combinations of those inputs directly on local hardware. That extended context lets a single embedding cover a substantial piece of real content, such as a recorded clip, a set of images, or a sequence of video frames, rather than only a short text snippet. Overall, EmbeddingGemma 2 brings robust embedding capabilities directly to edge hardware. Generating embeddings locally helps ensure data privacy, reduces pipeline latency, and empowers developers to build cross-modal search and retrieval that works entirely offline. Because the model is built on the Gemma 4 architecture and shares its text tokenizer and audio encoder, developers can run EmbeddingGemma 2 together with generative models such as Gemma 4 in a unified pipeline with a lower combined total memory footprint. That pairing enables on-device RAG pipelines that understand complex multimodal data, for example using EmbeddingGemma 2 for local file retrieval and Gemma 4 for contextual reasoning. The benefits described by Google follow directly from running embeddings on the device: data privacy, since content does not need to be sent to a cloud service; reduced pipeline latency; and offline operation, so search and retrieval keep working without a network connection. The modular parameter counts and MRL truncation reduce memory and storage requirements, letting developers index more content within a device's constraints. The quality gains in code, vision, and audio mean cross-modal search is practical at this size, and the extended context window means embeddings can represent longer stretches of audio, images, or video rather than only short text. Google demonstrates several concrete use cases. In Google AI Edge Gallery's Instant Media Search, users find the top matches in a media library using text or an image, ranked by semantic similarity. Video Moments Finder locates specific moments in video using text or audio queries. The Google AI Edge Foresight app pairs EmbeddingGemma 2 for local file retrieval with Gemma 4 for contextual reasoning. Developers can also create real-time decision engines that leverage multimodal context for classification, routing, and predictive capabilities via the MediaPipe Decision Task API. Because of the code improvement, the model is well suited for local codebase indexing, semantic code search, and coding agent retrieval. The launch example also describes finding a specific video clip from a voice memo, or searching through hours of audio recordings based on a text query, all handled by a single natively multimodal model. EmbeddingGemma 2 is aimed at developers building on-device and edge applications, including those working on search, retrieval, and RAG experiences. Getting started is supported by model weights on Hugging Face and Kaggle, with Gemini Enterprise Agent Platform Model Garden availability coming soon, and LiteRT Community on Hugging Face for models optimized for on-device use. For on-device deployment, developers can use Google AI Edge MediaPipe for turnkey embedding, retrieval, and decision tasks, or LiteRT for custom model integration; browser builds are possible with transformers.js or WebGPU. The model can be served with transformers, sentence-transformers, MLX, vLLM, llama.cpp, SGLang, Ollama, and LMStudio, embedding vectors can be stored with Qdrant, and Unsloth provides guidance on fine-tuning. Google also points to a developer guide, documentation, and specific inference and fine-tuning guides. The model is released under the commercially permissive Apache 2.0 license. EmbeddingGemma 2 continues the EmbeddingGemma line by moving lightweight embeddings beyond text. It unifies text, code, images, video, and audio in one shared embedding space, stays small enough to run on-device, and stays flexible through modular encoders and truncatable vectors. For developers who need privacy-first, low-latency search and retrieval that works offline, it provides a single multimodal model for cross-modal search, on-device RAG, and real-time decision tasks.