Google made EmbeddingGemma 2’s weights available Tuesday, expanding the model beyond text to code, images, video and audio. Google says the 740-million-parameter model puts those inputs into a shared embedding space for search and retrieval. Its image, audio and video benchmark results come from Google’s own evaluations; no independent test of those results was identified.
The weights are available on Hugging Face and Kaggle under the Apache 2.0 license, which permits commercial use, according to Google DeepMind’s announcement. Google describes the model as suitable for local use, though deployment depends on hardware, software integration and which components an application uses. A Google Model Garden listing was still forthcoming.
Embedding models turn inputs into numerical vectors that can be compared to find related items. Google says EmbeddingGemma 2 is designed to support searches across formats, such as using text or audio to find a moment in a video. Those are proposed uses, not independently verified results.
What does EmbeddingGemma 2 add beyond text?
Google introduced the first EmbeddingGemma as a text-embedding model in September 2025. The new version supports code alongside text, and adds optional vision and audio encoders for image, video and audio inputs. That could let developers build retrieval systems that search across formats rather than matching only text with text.
The model card lists a 270-million-parameter text component, a 170-million-parameter vision encoder and a 300-million-parameter audio encoder. Google says developers can leave out encoders they do not need, rather than include every component in an application.
EmbeddingGemma 2 produces 768-dimensional vectors, with options to shorten them to 512, 256 or 128 dimensions. Shorter vectors require less storage, but Google’s results show a trade-off at the smallest size: its overall MMEB v2 score falls from 59.01 at full dimensions to 45.65 at 128 dimensions. Google reports that the 256-dimensional version retains most of the quality on its listed benchmarks.
Google says a single input can include up to 5.5 minutes of audio, 29 images or 58 video frames, within an 8,192-token context window. A token is a unit used to process text. These are stated input limits, not a guarantee that every application will handle the material equally well.
What do Google’s multimodal benchmark results show?
Google reports scores of 64.64 on MIEB lite and 57.28 on MMEB v2 Image; 50.67 on MMEB v2 Video; and 69.54 on MSEB Retrieval and 49.39 on MAEB for audio. The figures come from Google’s model card. No independent evaluator’s reproduction of its image, video or audio results was identified.
For code, Google reports a score of 78.68 on MTEB Code v1, compared with 68.76 for the first EmbeddingGemma. The model card also lists a score of 61.36 on MTEB multilingual v2. These figures describe results on the tests Google reports; they do not establish how the model will perform on a particular organization’s collection of documents and media.
Qdrant, a vector database company, says it used early access to test storage configurations across five text-retrieval datasets. Its evaluation addresses the storage costs of reducing vector size, not the quality of image, audio or video retrieval. Independent, reproducible tests of EmbeddingGemma 2 across those modalities, including comparisons with named alternatives, were not identified.
Free newsletter
Get the morning briefing
Start each day with the AI stories that matter and why — a short, free email from our newsroom.
Can developers run EmbeddingGemma 2 locally?
Google says the weights are available on Hugging Face and Kaggle under Apache 2.0. Developers can download the weights rather than rely solely on an inference service hosted by Google. That makes local deployment an option, but does not establish that the full model will run on every device.
Google reports quantized active memory use of about 191 megabytes for the text-only weights and about 567 megabytes for the full multimodal model on a Google Pixel 11 Pro. Those are Google’s measurements, not independent tests on the device. The difference illustrates why a developer may choose to omit encoders an application does not need.
Local processing could let an application search media without sending indexed material to a remote service, as Google proposes. The model itself does not guarantee privacy or ensure that an application keeps all data on-device; those outcomes depend on how a product is built and operated.
Simon Willison has pointed to a broader reason developers may value open embedding weights: relying on a hosted provider can mean depending on its continued service or recalculating stored vectors if a model changes, adding work and cost. That is a concern about hosted embedding services, not an evaluation of EmbeddingGemma 2’s performance, as Willison discussed.
The weights are available, but no independent evidence of the model’s image, audio and video retrieval was identified. Tests across varied collections could show whether Google’s benchmark results transfer to practical mixed-media search. Google has not given a date for adding the model to Model Garden. Qdrant’s evaluation concerns storage and text retrieval, not multimodal performance.