The Pulse
Google Releases 740M-Parameter EmbeddingGemma 2 for Edge AI
Google launched EmbeddingGemma 2, an open-weight model that maps text, images, video and audio into a shared embedding space for on-device search and retrieval. Its modular encoders let developers choose a smaller footprint, while Google sa
AI.info Team · Written with AI ·
Type “Katze” into Google’s new media-search demo and photos of cats in a phone’s local collection appear; add “schläft auf Tastatur” (“sleeping on keyboard”) and the results narrow to the matching image. Google presented that search-as-you-type example on October 6, when it announced EmbeddingGemma 2, a 740-million-parameter model built to search across text, images, video and audio on consumer devices.
The model converts different kinds of input into vectors in a shared 768-dimensional space, allowing a text query to retrieve an image or a video moment without first turning the media into captions or transcripts. Google says its open-weight release is intended for local search and retrieval, reducing the need to chain separate models for tasks such as image captioning, speech recognition and text embeddings.
One model, four input types
EmbeddingGemma 2 is modular rather than a single fixed-size package. Its text-and-code components total 270 million parameters; developers can add a 170-million-parameter vision encoder, a 300-million-parameter audio encoder, or both. The full configuration reaches 740 million parameters, while unused encoders can be left unloaded.
Google’s model card lists an 8,192-token context window and support for more than 100 languages. It also reports a roughly 14% improvement over EmbeddingGemma 1 on code tasks. For applications that store large search indexes, developers can shorten output vectors from 768 dimensions to 512, 256 or 128; Google’s evaluations show that shorter vectors cut storage needs, with some reduction in benchmark scores.
Search stays on the device
Google’s AI Edge Gallery adds two demonstrations: Instant Media Search finds local photos and videos using natural-language queries or example images, while Video Moments Finder locates visual moments in recordings. The announcement says the gallery creates embeddings on the phone and keeps its media index in a local SQLite database. The demos are available through the gallery app for Android and iOS.
Google also introduced AI Edge Foresight for Mac, an experimental meeting companion that processes audio and files locally. The company says it can enrich shorthand notes with meeting context and retrieve information from transcripts, notes, images and documents, including when the computer is offline.
Android integration is still pending
Developers can use Google’s MediaPipe Tasks for cross-platform embedding and retrieval, or LiteRT for more direct control over execution. Google says its Universal Embedder handles image preprocessing and returns vectors, while its Semantic Retriever performs on-device approximate-nearest-neighbor searches. A MediaPipe Decision task can also compare inputs with labels for classification or routing without fine-tuning.
Google’s launch post reports vision-embedding latency of 37.3 milliseconds per image on a MacBook M5 Pro GPU, measured with a maximum budget of 70 vision tokens per image. That result describes one workload on one device, not a general speed guarantee. The post says Google will bring the model to Android through ML Kit in the coming weeks, with neural-processing-unit acceleration where available; that integration was not yet available at launch.
Open weights, with trade-offs to test
Google lists the model under the Apache 2.0 license and offers pre-quantized LiteRT bundles alongside the full weights. Its model card reports results across text, code, image, video and audio benchmarks, but those figures come from Google’s own evaluation and use the full-precision checkpoint. Developers choosing smaller or quantized configurations will need to test accuracy and speed on their intended hardware.
For now, the most concrete promise is local retrieval: find a photo with a sentence, locate a moment in a recording, or match meeting notes to a transcript without sending that search workload to a cloud service. Android app developers will have to wait for the announced ML Kit release to get Google’s managed integration.