EmbeddingGemma 2 Operational

EmbeddingGemma 2 is Google's open-source native multimodal embedding model. Based on the Gemma 4 architecture, it maps text, code, images, videos, and audio into a unified vector space for cross-modal semantic search. EmbeddingGemma 2 uses a modular design, with just 740M parameters across all modalities, an 8K-token context window, and approximately 567MB of memory usage on phones after quantization. EmbeddingGemma 2 is the strongest multimodal vector model under 1B parameters, improving code retrieval by nearly 10 points over the previous generation. It is suitable for on-device RAG, semantic search, photo album retrieval, and other scenarios.

Cross-modal semantic search: Maps text, code, images, videos, and audio into the same vector space, enabling any modality to be used to retrieve content in other modalities.
On-device local execution: With just 740M parameters across all modalities, it uses approximately 567MB of memory on the Pixel 11 Pro after quantization, enabling inference on phones and laptops without an internet connection.
Local RAG pipelines: Shares a tokenizer and audio encoder with Gemma 4. Together, they can build a multimodal retrieval-augmented generation system locally while protecting data privacy.
Semantic photo and media search: In Instant Media Search in Google AI Edge Gallery, users can enter text to search images in their phone's photo album or use the camera to photograph an object and find similar photos.
Video moment localization: Video Moments Finder supports text or audio queries and automatically identifies semantically matching segments in videos.
Real-time decision engine: Uses the MediaPipe Decision Task API for real-time classification, routing, and prediction. In demonstrations, a single decision took only 43 milliseconds, much faster than the 640 milliseconds required by an LLM.

Follow WeChat and reply “开源” to join the AI open-source project community
Unified multimodal embedding space: The model follows the technical approach of the Gemini Embedding series, encoding inputs from different modalities into the same high-dimensional vector space. Semantically similar content is therefore close together in vector space, while unrelated content is farther apart. Cross-modal retrieval becomes a unified vector nearest-neighbor search problem, eliminating the need to maintain a separate model for each modality.
Modular encoder-decoder architecture: Built on the Gemma 4 architecture, the model combines three components: a 270M-parameter text backbone, a 170M-parameter vision encoder, and a 300M-parameter audio encoder. The vision and audio modules can be loaded as needed, resulting in four deployment configurations: text only, text+image (440M), text+audio, and all modalities. Text tasks therefore avoid the memory overhead of the vision and audio modules.
Shared components with Gemma 4: EmbeddingGemma 2 and the generative model Gemma 4 share the same text tokenizer and audio encoder. They can be connected into a unified on-device RAG pipeline: EmbeddingGemma 2 encodes local multimodal files for indexing, while Gemma 4 performs contextual reasoning after retrieval. Shared components also reduce total memory usage during combined deployment.
Matryoshka representation learning and dynamic truncation: The model uses MRL training and outputs 768-dimensional vectors, which can be dynamically truncated to 512, 256, or 128 dimensions during inference. Truncating to 256 dimensions causes virtually no performance loss, while vector database storage can be reduced to as little as one-sixth of the original size.
5. Long context and long-input chunking: The model supports an 8K-token context window, four times that of the previous generation. A single request can process approximately 5.5 minutes of audio, 29 images, 58 video frames, or an interleaved combination of these inputs. Text inputs also support task prefixes, allowing the same text to generate different vector representations for different tasks.

Download the model weights: Obtain the model files for google/embeddinggemma-2 from Hugging Face or Kaggle. For an on-device optimized version, visit the LiteRT Community page.
Choose a runtime framework: Load the model directly with mainstream tools such as transformers, sentence-transformers, MLX, vLLM, llama.cpp, SGLang, Ollama, and LM Studio.
Write inference code: With sentence-transformers, for example, load the model, encode queries and documents using prompt_name="SearchQuery" and "Document" respectively, then call similarity() to calculate vector similarity.
Build a retrieval system: Store the generated embeddings in a vector database such as Qdrant to enable semantic search, codebase indexing, or multimodal retrieval.
Deploy on devices: On mobile, use Google AI Edge MediaPipe to directly integrate embedding, retrieval, and decision tasks, or use LiteRT for custom integration. In browsers, run the model with transformers.js / WebGPU.
Build a local RAG system: Use it with Gemma 4, with EmbeddingGemma 2 handling local file retrieval and Gemma 4 handling contextual reasoning. Their shared components reduce total memory usage.

Native unified multimodality: A single model covers text, code, images, videos, and audio, mapping all five modalities into the same vector space instead of combining separate models for each modality.
Extremely lightweight and suitable for on-device use: With just 740M parameters across all modalities, it uses approximately 567MB of memory on the Pixel 11 Pro after quantization, enabling offline multimodal inference on phones.
Best performance in its size class: It is the strongest multimodal embedding model under 1B parameters, even surpassing specialized models more than twice its size on some tasks.
Major gains in code capabilities: Code retrieval scores improve by nearly 10 points over the previous generation, making it highly suitable for local codebase indexing and coding-agent retrieval.
Modular on-demand loading: The vision and audio encoders can be detached independently. Text tasks require only 270M parameters, avoiding memory costs for unnecessary modalities.
Low storage cost: Based on Matryoshka representation learning, 768-dimensional vectors can be truncated to 128 dimensions, reducing vector database storage to as little as one-sixth of the original size.
Long-context support: The 8K-token context window is four times that of the previous generation, allowing a single request to process approximately 5.5 minutes of audio or 29 images.
Native synergy with Gemma 4: The models share a tokenizer and audio encoder, reducing total memory usage when combined to build a local RAG system.

Official website: https://blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2/
Hugging Face model repository: https://huggingface.co/google/embeddinggemma-2
Technical paper: https://ai.google.dev/gemma/docs/embeddinggemma/model_card_2

On-device multimodal semantic search: Search a local media library from a phone's photo album using text or photos taken with the camera, entirely offline and without uploading private data.
Video content retrieval and moment localization: Use text or audio queries to quickly locate target segments in surveillance footage, online courses, or meeting recordings, such as scenes of “sea turtles feeding,” without manually searching frame by frame.
Local RAG knowledge bases: Build an offline RAG system with Gemma 4, uniformly indexing local documents, images, and audio. This is suitable for enterprise intranets, classified environments, and other scenarios where cloud access is not permitted.
Semantic codebase indexing and agent retrieval: Build a vector index for a local codebase to support semantic code search for coding agents. Code retrieval scores improve by nearly 10 points over the previous generation.