What Runs on Your Device: Mac, Browser, and iPhone Compared

"On-Device" Means Different Things in Different Places

"On-device AI" gets used as a single marketing phrase, but what it actually means depends heavily on where you're running it. A laptop with a real CPU and GPU can run a language model locally. A phone usually can't, not for anything beyond the smallest tasks. A browser tab sits somewhere in between: capable of running speech recognition locally, but not built to run a full summarization model the way a desktop app can.

Grafite treats these as genuinely different situations rather than pretending they're the same. Here's what actually runs where, plainly, including the parts that aren't flattering.

Mac Desktop App: Transcription and Summaries Can Both Run Locally

On the Mac app, both transcription and summarization can run entirely on your machine, using a local model you choose in Settings, without your audio or transcript leaving the device for either step. This is the deepest local option Grafite offers, and it's only possible because a desktop app can run real background processes with real memory and CPU budgets that a browser tab or a phone can't.

It comes with a hardware floor. Running the full local pipeline, transcription and summarization together, at Grafite's recommended quality needs a machine with at least 8 GB of RAM and 4 CPU cores. Below that, the app can still run a lighter on-device transcription model, but the fuller local experience needs the hardware to back it up. If your Mac doesn't clear that bar, Grafite falls back to cloud processing and tells you why, rather than silently sending your audio off the device without saying so.

You also get a real choice about where your notes live: on the device only, or synced through your account. That's a separate decision from where the AI processing happens, but the two are related. If you want your meeting data to never touch Grafite's servers at all, keeping notes local plus running transcription and summarization locally gets you close to that, for the parts of the app that don't need an internet connection to function.

The Honest Caveat

"On-device" is not the same claim as "nothing ever leaves your device." Even with local transcription and local summaries and local notes, some things still need the internet and still touch Grafite's servers: signing in, calendar sync, anything you explicitly share, and cross-device Ask all depend on an account and a network connection by design. Local mode narrows what leaves your machine. It doesn't eliminate every network call the app makes.

In the Browser: Transcription Locally, Summaries in the Cloud

Open Grafite in a browser tab, no app install, and you still get on-device transcription. Speech-to-text can run directly in the tab using a local model, which means the audio itself doesn't have to leave your browser to become a transcript.

Summarization is where the browser draws a line the desktop app doesn't. Turning a transcript into a structured summary, extracting next steps, all of that runs in the cloud when you're in a browser, not locally. A browser tab doesn't have the same background process model or resource budget a native app does, so the heavier generation step happens server-side.

Practically: if you want transcription without an install, the browser gets you there. If you want the whole pipeline, transcription and summarization, running locally, that's the Mac app's job specifically.

iPhone: Cloud Today, No Exceptions

On the iPhone app, transcription and summaries both run in the cloud. There's no on-device option on iOS right now, for either step. If privacy from cloud processing matters more to you than mobile convenience, the iPhone app currently isn't the way to get it; the Mac app is.

This is worth stating without softening it, because it would be easy to let "on-device is available" bleed across platforms in marketing copy when it isn't actually true everywhere. It isn't. iPhone is cloud-only, full stop, today.

What "Local Only" Actually Buys You on the Desktop App

Choosing local processing on the Mac app is a meaningful decision, not a cosmetic setting. When both transcription and summarization run on-device and your notes are set to stay on the device, the meeting audio, the transcript, and the generated summary all exist only on that machine unless you later choose to share or sync them. That's a real difference from a cloud pipeline, where the audio and transcript pass through a server even if you trust the company running it.

The trade for that is that you're relying on your own hardware instead of a data center's. A capable Mac handles this well. An older or lower-spec machine will feel it, both in processing time and in the fan spinning up during a long call. Grafite's hardware gate exists precisely so the app doesn't quietly degrade into an unusably slow local pipeline on a machine that can't really support it; below the cutoff, it defaults to cloud and explains why, rather than leaving you to figure out why summaries are taking five minutes.

Why the Trade-Offs Exist

Running a model locally isn't free, even when the hardware can technically do it. A few honest trade-offs worth knowing before you pick a mode:

Speed. Cloud models, especially the higher-tier ones, are frequently faster than what a laptop can run locally, particularly for longer meetings. Local processing on modest hardware can take noticeably longer to produce a summary.

Battery and heat. Running transcription and a language model locally uses real CPU and GPU cycles. On a laptop, that means more fan noise and faster battery drain during a long meeting than a lightweight cloud round-trip would cause.

Accuracy versus the premium cloud engine. Grafite's cloud transcription, on Professional plans, uses a higher-accuracy engine than what runs on-device today. If transcript accuracy on a noisy call matters more to you than keeping audio off the network, cloud is the stronger choice for that specific job.

Speaker labels are cloud-only. Telling apart who said what in a multi-person call is a cloud, Professional-plan feature. It isn't available in the on-device pipeline, on any platform. If your meetings regularly need per-speaker attribution, that pulls you toward cloud transcription regardless of platform. We went deep on why that specific feature is technically hard to do well in why speaker labels are hard.

Picking a Mode That Fits How You Actually Work

None of this is really about which mode is "better" in the abstract. It's about which trade-off fits a given meeting. A sensitive internal call on a capable Mac is a reasonable case for full local processing, even if it's a little slower. A quick call from a laptop that's already low on battery, or a call where you need clean speaker labels for a transcript you'll share afterward, is a reasonable case for cloud. On the iPhone, right now, there isn't a choice to make. It's cloud, and that's fine for a lot of everyday calls, worth knowing plainly rather than assuming otherwise.

The honest version of "on-device AI" is not a single toggle that means the same thing everywhere. It's a set of real trade-offs that differ by platform, and knowing which is true on your device is more useful than a blanket claim would be. For more on why keeping meeting data personal matters in the first place, see why meeting recordings should be personal, and if you dictate outside of meetings too, dictating across Slack, docs, and email covers the same local-versus-cloud choice for that workflow.

See Grafite's on-device options for yourself, download the Mac app to try local transcription and summaries, or use the browser for on-device transcription without installing anything.