Summary
Google has launched EmbeddingGemma 2, a 740-million-parameter model that maps text, code, images, audio and video into a shared embedding space. The company says it is designed for on-device search and retrieval, with modular encoders and adjustable vector sizes.
Google has launched EmbeddingGemma 2, a 740-million-parameter model that creates embeddings for text, code, images, video and audio. The company says it maps these different kinds of content into a shared embedding space, enabling searches that can cross between them, such as using a voice memo to locate a moment in a video.
Google describes the model as built for on-device use and says it is released under the commercially permissive Apache 2.0 licence. Its intended uses include local media search, codebase indexing and retrieval-augmented generation (RAG), where a system retrieves relevant information to provide context to a generative model.
Cross-modal search with configurable model size
An embedding represents information as a list of numbers, or vector. Systems can compare these vectors to find items that are similar in meaning, rather than relying only on exact keyword matches. When different types of content share an embedding space, a text or audio query can be compared with images, video or other media. Google gives examples including searching hours of recordings with a text query and finding a video clip using a voice memo.
The model is modular. Google says text-only workloads can use as little as 270 million parameters, with optional vision and audio encoders of 170 million and 300 million parameters respectively for multimodal support. It uses Matryoshka Representation Learning, a technique that allows developers to shorten output vectors: EmbeddingGemma 2’s 768-dimensional vectors can be reduced to 512, 256 or 128 dimensions. Google says the shortest setting can reduce storage for local vector databases and memory use by up to six times.
The model has an 8,000-token context window, four times the size of EmbeddingGemma 1’s, according to Google. The company says it can handle up to 5.5 minutes of audio, 29 images, 58 video frames, or interleaved combinations of these in a single context.
Performance claims and on-device requirements
Google reports a score of 78.68 on MTEB Code, compared with 68.76 for EmbeddingGemma, a 9.92-point increase. It says the new model maintains the earlier model’s multilingual text performance and achieves leading results among multimodal embedding models with fewer than one billion parameters on benchmarks including MTEB Code and the Massive Audio Embedding Benchmark. The announcement points developers to the model card for full evaluation metrics.
For local use, Google reports that, with quantization, EmbeddingGemma 2 uses about 191 MB of active RAM for text-only weights and about 567 MB for the full multimodal model on a Google Pixel 11 Pro. Quantization reduces the numerical precision used to store model weights, generally lowering memory requirements. Google says local embedding generation can reduce processing latency, support privacy-focused retrieval and allow search to work offline.
The company also points to Google AI Edge Gallery demonstrations for media search and video moment finding, and to its MediaPipe Decision Task API for multimodal classification and routing. These examples show the intended role of the model: turning locally stored mixed-media content into material that applications can search and organise without sending every query through a remote service.