Google releases free AI Edge Foresight, a Mac note-taking app that transcribes meetings fully offline
Google has released AI Edge Foresight, a free experimental macOS app that transcribes meetings and audio files entirely on-device. It is optimized for Apple Silicon Macs and uses Google's EmbeddingGemma 2 model, according to The Verge. Google says audio, files, and notes never leave the user's computer.
Google has released AI Edge Foresight, a free, experimental note-taking app for macOS that transcribes meetings and audio files entirely offline. The app runs on Google's on-device EmbeddingGemma 2 model, according to The Verge, which cited earlier reporting by TechCrunch.
What the app does
- Offline transcription: Foresight transcribes meetings and audio files locally, with no cloud connection required.
- Note enhancement: Users type shorthand bullet points during a meeting. Google says the app turns them into "polished notes" based on the transcript.
- Built-in assistant: An assistant answers questions about what was said in a meeting.
- Local file context: Users can connect local files, which the assistant can reference when answering questions.
- Privacy: Google says files, meeting audio, and notes "never leave your computer." This is a company claim. No independent verification or security audit has been reported.
Availability and specs
| Detail | Status |
|---|---|
| Platform | macOS only |
| Hardware | Optimized for Apple Silicon Macs |
| Price | Free |
| Status | Experimental |
| Model | EmbeddingGemma 2 (per The Verge) |
| Parameter count | Not disclosed |
| Context window | Not disclosed |
| Benchmarks | None published |
| Windows, Linux, mobile support | Not announced |
The app joins a growing category of AI meeting-note tools. The Verge compares it to Granola and Wispr Flow. Google's differentiator is on-device processing.
Open questions
The source reporting names EmbeddingGemma 2 as the underlying model but does not explain how it handles speech transcription. Embedding models typically convert text into vectors for retrieval and similarity search, rather than generating speech-to-text or summaries. Foresight may combine several on-device components, such as a speech recognition model, a generative model for note enhancement, and an embedding model for retrieving local files. Google has not detailed that architecture in the material reviewed. It has also not published transcription accuracy figures, supported languages, minimum RAM requirements, or storage footprint.
What this means
For teams that handle sensitive meetings, such as legal, healthcare, finance, or unreleased-product discussions, the main barrier to AI note-takers has been sending audio to third-party servers. A free, local-only option from Google addresses that objection directly, provided the privacy claims hold up under scrutiny.
The launch also shows Google using its AI Edge tooling to ship a complete consumer-facing application rather than only developer components. The Apple Silicon-first focus suggests the app depends on local compute and memory bandwidth that current Macs provide.
The "experimental" label matters. Without published accuracy benchmarks, language coverage, or hardware requirements, it is too early to say how Foresight compares with cloud-based competitors, which can run larger models. The practical test will be transcription quality and summary reliability on real meetings with multiple speakers and technical vocabulary.
This is a product launch built on an existing model, not a new model release.
Related Articles
Tencent opens WorkBuddy Computer Use beta for Windows and macOS apps that lack APIs
Tencent announced a public beta of WorkBuddy's Computer Use feature on November 20. It lets the office agent operate desktop applications that have no API, connector or browser version, on Windows and macOS, and requires WorkBuddy client 5.11.0 or later.
Tunix agentic GRPO lifts Gemma 4 E2B from 9.4% to 95% win rate in Plane Strike
Google's developer blog details how Tunix's agentic GRPO pipeline trained Gemma 4 E2B to play Plane Strike with tool calls. The win rate on held-out boards rose from 9.4% to 95% after 100 steps on a single Cloud TPU v6e-8 host.
Google previews Tool Retrieval in Gemini Interactions API to load tool schemas only when needed
Google announced Tool Retrieval on November 19, 2026, available in preview in the Gemini Interactions API. Agents can keep large tool libraries on standby and load full parameter schemas only when the current task needs them.
Google makes Genie 4 available in Gemini API and AI Studio with 720p, 24 fps interactive worlds
Google made Genie 4 available through the Gemini API and Google AI Studio on November 19, 2026. The model generates continuous 720p video at 24 frames per second with synchronized native audio, and worlds respond to keyboard controls and live text prompts.
Comments
Loading...