Google AI Edge Foresight is not a Granola competitor. It is a hardware demo wearing a productivity app costume, and that makes it far more useful to those of us who build bots than any meeting summarizer could ever be.
The mainstream read on this release, which TechCrunch covered on October 8, 2026, goes something like: big company notices small company has nice product, big company clones nice product, we all sigh about platform risk. Fair enough as a narrative. But I have been shipping on-device inference into assistant projects for a while now, and when I look at Foresight I do not see a note-taking app. I see Google publishing a reference implementation for a problem I have been solving badly with duct tape.
What Was Actually Shipped
The facts are thin, and I will not pretend otherwise. Foresight is a Mac app. It captures meeting notes. It runs offline. It uses EmbeddingGemma 2, an on-device model with 740 million parameters. Google has not disclosed the full feature set, the supported platforms, or how it plans to make money from this. TechCrunch notes that the same team released an AI-powered dictation tool back in April that also worked with local models.
That is the whole verified picture. Everything else floating around is speculation, and some of it is people on social media joking about their “Granola Bro.” Let the hype cycle handle that part.
Why the Parameter Count Is the Story
Seven hundred and forty million parameters is small. Not small in a cute, toy-project sense, but small in a this-fits-in-memory-alongside-your-actual-application sense. For anyone who has tried to run a local assistant and watched a 7B model eat the entire machine, that number should register as a design statement.
And it is an embedding model, which tells you something about the architecture. Embeddings are how you turn messy human speech into something searchable and groupable without needing a giant generative model to reason over every token. If you are building a bot that needs to recall what a user said three weeks ago, you do not need a frontier model. You need good vectors and a sensible retrieval layer.
Google appears to be betting that the useful parts of a meeting assistant are mostly retrieval problems dressed up as generation problems. I think that bet is correct, and I think most of us building in this space have been overspending on the generation half.
What This Changes for Bot Builders
Here is the practical angle for anyone working on assistants, agents, or anything that touches private user data:
- Local-first just got a vendor-blessed blueprint. When Google ships a Mac app running a small on-device model, that is implicit documentation for the deployment path you have been guessing at.
- Embedding-first architectures are getting a push. If EmbeddingGemma 2 is good enough to anchor a shipped Google product, it is probably good enough for your retrieval layer.
- Offline is a feature you can now promise. Convincing a user that their meeting audio never leaves the laptop used to require a lot of hand-waving. It gets easier when a major platform makes the same claim.
- The privacy-first competitors are validated, not crushed. Projects like Char, the open-source local-first meeting notepad formerly called Hyprnote, built on the premise that on-device data handling matters. Google agreeing with them is a tailwind.
The Part I Am Skeptical About
No disclosed business model is a yellow flag. Google has a history with experimental apps that behave like experiments. If you are thinking about building on top of Foresight rather than learning from it, that undisclosed model should slow you down. Build against the pattern, not the product.
No disclosed platform support is a second yellow flag. A Mac-only release from Google reads as a pilot. That is fine for a pilot, less fine if you were hoping to ship cross-platform on the same foundation next quarter.
What I Would Build Next
If I were starting a local assistant project this week, I would stop reaching for a large model as the default and instead structure the thing around a small embedding model plus a solid local store. Capture, embed, index, retrieve. Add generation only where it earns its cost. Foresight suggests Google’s own team arrived at a similar conclusion after shipping dictation first and note-taking second, which is a sensible order of operations.
The Granola comparison will drive the headlines because comparisons are easy. The more interesting question is whether small on-device models are quietly becoming the default substrate for assistant work rather than a compromise you make when the cloud is unavailable. A 740 million parameter model in a shipped Google product is a reasonable piece of evidence for the former.
I will be pulling EmbeddingGemma 2 into a test use this week. Expect a build log.
🕒 Published: