Google released a Mac note-taker that runs the model on your laptop

Taking notes about a confidential meeting with a cloud note-taker means the audio is sent somewhere, kept for some time, and transcribed by someone else's model. Google released a Mac app called Google AI Edge Foresight that does the same job without any of that: the audio, the transcript, the files it reads, and the answers it writes all stay on the Apple Silicon machine it runs on.

The app is a free download for macOS on ai.google.dev/edge/foresight. It runs on EmbeddingGemma 2, Google's 740 million parameter on-device embedding model, together with Gemma 4 as the chat model. On its FAQ page, Google says that "all AI inference runs 100% locally on your device using on-device Gemma models" and that "no personal files, recordings, or transcripts are ever transmitted to external cloud servers."

What the app does

Foresight listens to a meeting in the background, from either a video call or an in-person conversation, and transcribes the audio locally. A split screen lets the user type shorthand bullets on one side while the transcript builds on the other. After the meeting, Foresight rewrites the shorthand into polished notes using the transcript as context.

Users can point the app at local folders, Google Drive, PDFs, Microsoft Office files, Markdown, plain text and web bookmarks. These become a knowledge base the app can read during a meeting. A Live Assistance mode listens for questions the meeting raises and surfaces the matching passage from the knowledge base with a citation.

The app also runs a chat window where a user can ask questions in natural language and get answers drawn from the meeting and the imported files, without a network connection.

Where this sits

Foresight is a competitor to Granola, Fathom and the other paid note-takers that run in the cloud. The TechCrunch report by Ivan Mehta lists the recent moves in the space: Wispr and Calendly released their own note-takers, and Superhuman, formerly Grammarly, acquired Fathom.

Google says the Gemma model and the vector embeddings engine run on the user's hardware with no internet connection required, so the app also works on a plane or in an air-gapped environment. The hardware requirement, in Google's own words, is Apple Silicon with "lightweight Gemma quantization and memory-efficient vector kernels."

The team that shipped Foresight is the same group behind Google's existing on-device dictation tool. TechCrunch reads Foresight as a showcase for what the Gemma series can do without a cloud rather than a product push on its own.

For a team building AI features, the move worth watching is Google releasing a shipped Mac app that runs end to end on a 740 million parameter embedding model and the Gemma 4 chat model. The embedding model, Gemma 4, and the retrieval pipeline that joins them come from Google's existing public releases, and Foresight is one worked example of how far they go when run locally. The pattern (local embedding plus a small local chat model plus a vector index that stays on the device) is the same one a product team would use for any feature where data has to stay inside the customer's machine.

Source

Google AI Edge Foresight, Google · Google releases a new local-first Granola competitor, Ivan Mehta, TechCrunch