Google demonstrates EmbeddingGemma 2 in Foresight for Mac

Google says the downloadable app expands shorthand with meeting transcripts and retrieves project information from local files and Google Drive.

Google says the downloadable app expands shorthand with meeting transcripts and retrieves project information from local files and Google Drive.
Google has made AI Edge Foresight available to download for Mac as a demonstration of EmbeddingGemma 2, according to 9to5Google’s October 6 report. The app puts the model Google announced that day into a notetaking workflow: using recorded discussion to expand short notes and retrieving information from project materials.
According to Google’s description reported by 9to5Google, Foresight listens in the background during video or in-person meetings while users write shorthand bullet points. The app then uses the meeting transcript to turn those bullets into fuller notes. Google says the meeting audio and transcript are processed entirely on the user’s device rather than in the cloud.
Google also says users can point Foresight to project folders, reference materials, diagrams and their calendar. With support for local files and Google Drive, users can ask questions to retrieve information from those materials. Google describes the chat feature as offering conversational retrieval and summaries of the user’s information.
The note-expansion and retrieval capabilities are Google’s claims, not independent accuracy findings. The on-device processing claim concerns meeting audio and transcripts; how Foresight handles Google Drive data is not yet clear.
EmbeddingGemma 2 is a superintelligence (SI) model built on Gemma 4, which Google released in April. SiliconANGLE reports that it is more than twice the size of the original model. Google describes two separate ways developers can reduce its resource needs: omit unused media components to save memory, and shorten its embeddings—lists of numbers that represent content for search—to save storage.
Most of the model’s growth comes from its vision and audio encoders, the components that convert those inputs into numerical representations. SiliconANGLE reports that text-only apps can omit them and use the text core instead.
Google’s memory figures use quantization, which reduces the numerical precision of model values to save memory. Parameters are the learned numerical values that define a model. Both measurements below concern a Pixel 11 Pro; they are Google’s own results and do not establish memory use or performance on other hardware.
| Configuration | Parameters | Active RAM |
|---|---|---|
| Text core | 270 million | About 191 MB |
| Full model for text, images, video and audio | 740 million | About 567 MB |
Running the model is only one resource demand. An app also stores its outputs in a vector database, a search index containing embeddings. Every indexed photo, clip or document adds another embedding to that database.
A training method called Matryoshka Representation Learning lets developers shorten these lists of numbers. The storage saving described below applies to stored embeddings, not to the model’s RAM use or all the storage an app needs.
| Numbers per embedding | Reported role or tradeoff |
|---|---|
| 768 | Full-length reference |
| 256 | Google claims about 95% of full image, video and speech retrieval quality |
| 128 | Up to sixfold less embedding storage than at full length |
Shorter embeddings can reduce how well searches find relevant content. The quality claim comes from Google’s developer guide, as reported by SiliconANGLE, and applies to 256 dimensions—not the shortest 128-dimensional option. It is a vendor claim rather than an independently verified result and does not establish the quality developers will get on their own data.
9to5Google’s October 6 report points to a Foresight download for Mac and to EmbeddingGemma 2 demonstrations in Google AI Edge Gallery on Android and iOS. The model is available under the Apache 2.0 licence.

One bill would prohibit federal agencies from using license-plate readers and restrict funding to other governments that use them. Another would set search approvals and data-retention limits.
4 min

The Document Foundation says built-in SI does not yet meet its requirements for keeping data on users’ devices and avoiding dependence on one provider.
2 min

The recovery follows a spring selloff over fears that businesses could build their own applications more cheaply.
2 min