Remember back in April, when Google quietly shipped an AI dictation tool that ran on local models? It didn’t get a keynote. It didn’t get a hype cycle. It showed up, did a narrow job on-device, and most people who review tools for a living filed it under “interesting, probably abandoned by Q3.”
That same team just shipped again. The new thing is called Google AI Edge Foresight, it’s a Mac app, and it’s going directly after Granola — the AI note-taker that a lot of us have been quietly paying for because it does one thing well. Foresight captures meeting notes offline using EmbeddingGemma 2, an on-device model with 740 million parameters.
That’s the whole verified picture right now. Google hasn’t disclosed the full feature set, which platforms beyond Mac are coming, or how it plans to make money. So let me be clear about what this piece is: not a review. I haven’t run it against my own meeting load yet. This is me telling you what signals I’m reading and which questions I’d want answered before switching anything.
Why the local-first part actually matters here
Meeting notes are the worst possible category for cloud processing, and everyone knows it. Your standups contain roadmap details. Your 1:1s contain personnel discussions. Your client calls contain things covered by an NDA you probably signed without reading. Every cloud note-taker asks you to trust a vendor’s retention policy with the most sensitive unstructured text your company produces.
That’s why the local-first corner of this category has been getting crowded. Char, formerly Hyprnote, rebranded in early 2026 as an open-source local-first meeting notepad built on the same keep-it-on-device philosophy. Granola itself became popular partly on vibes of restraint. The demand is real, and it’s not coming from privacy absolutists. It’s coming from people who want to record a meeting without filing a vendor security review first.
Google showing up in that corner is the genuinely surprising part. Google’s business model historically runs through its servers. An app that processes your meetings on your laptop and never ships the transcript anywhere is a strange fit for that machine. Either this is a sincere bet that on-device inference is where the next generation of productivity tools live, or it’s an experiment from a team with enough internal runway to try something off-brand. The April dictation tool suggests the former — same team, same approach, second shipped product.
The 740 million parameter question
EmbeddingGemma 2 at 740M parameters is small by frontier-model standards, and that’s the point. It fits on a laptop, runs without a network call, and doesn’t melt your battery during a two-hour planning session. The tradeoff is capability. An embedding model of that size is good at understanding and organizing text. It is not GPT-4 writing you a beautifully structured summary with action items inferred from tone.
This is where I’d expect the real comparison against Granola to land. Granola’s appeal has never been transcription accuracy; it’s the quality of what comes out the other end. If Foresight’s on-device processing produces notes that are merely organized rather than genuinely useful, privacy won’t save it. Most people trade privacy for output quality every single day without thinking about it.
So the honest framing is a question, not a verdict: can a 740M-parameter model on your own hardware get close enough to cloud-quality summaries that the privacy benefit tips the decision? I don’t know yet. Nobody outside Google does.
What I’d want to know before switching
- Summary quality against real meetings. Not a demo call. A messy five-person discussion with crosstalk and someone’s dog.
- The business model. Google hasn’t said. Free-forever on-device tools from large companies have a specific historical pattern, and it usually ends with a sunset notice.
- Platform support. Mac-only today. If half your team runs Windows, this isn’t a team tool yet.
- Export and lock-in. Local-first is only meaningful if the data is actually yours in a format you can move.
- Whether the team survives. Two shipped products in roughly six months is a solid sign. It’s not a guarantee.
My read
If you’re currently paying for Granola and it works, don’t churn on an announcement. Switching note-takers mid-quarter costs you more in friction than you’ll gain in principle.
If you’ve been avoiding AI note-takers specifically because you don’t want meeting transcripts sitting on someone else’s infrastructure, this is the first entry in that category with serious engineering weight behind it. Worth a trial on low-stakes meetings.
And if you’re building in this space, the signal is loud: on-device inference just got a competitor with Google’s distribution. The privacy pitch alone stopped being a moat this week.
I’ll run Foresight against my actual calendar and report back with numbers instead of speculation. That piece will be more useful than this one.
🕒 Published: