Hexagone AI runs the AI models on your own machine, so the answer depends on your machine. Here is what we require, how long a typical file takes on the laptops we test with, and what makes it faster.
Describe your computer and the job, and get an estimated time with a plain verdict. It uses the measured runs further down this page as its anchors.
25 s
10 pages, fast mode, digital
Likely 20 s to 40 s · 2 s per page
You will rarely wait. This is the experience the app is designed around.
An estimate from the measured runs further down, not a promise. The app times your actual machine at first launch.
The minimum column runs the app. The recommended column is where it feels quick, and where Accurate mode is comfortable to use every day.
| Component | Minimum | Recommended |
|---|---|---|
| Chip | Apple Silicon (M1 or later) | M2 or later, any Pro or Max chip |
| System | macOS 13 Ventura | macOS 14 or later |
| Memory | 8 GB | 16 GB or more |
| Free disk space | 3 GB | 5 GB |
| Component | Minimum | Recommended |
|---|---|---|
| Processor | 64-bit Intel or AMD, 4 cores | Intel Core 12th gen or Ryzen 5000 or later, 6 cores or more |
| System | Windows 10 64-bit with WebView2 | Windows 11 |
| Memory | 8 GB | 16 GB in two sticks (dual channel) |
| Free disk space | 3 GB | 5 GB |
| Graphics card | None required | Discrete NVIDIA, Radeon RX or Intel Arc for Accurate mode |
Intel Macs, Windows on ARM and Linux desktops are not supported by the desktop app. Teams on those machines can use the hosted version at app.hexagone.ai or, on the Enterprise plan, an on-premise deployment. See the Enterprise plan.
A 10-page Word document of about 30,000 characters with roughly 25 pieces of personal data per page, as digital text. Times are for the whole file, from drop to protected copy. A scanned PDF of the same 10 pages adds the OCR column.
| Your machine | Fast mode | Accurate mode | If scanned, add |
|---|---|---|---|
| Apple Silicon Mac, 16 GB or moreAny MacBook Pro, or an Air, mini or iMac with 16 GB | 15 to 30 sestimated | 4 to 6 minestimated | + 10 to 20 sestimated |
| Apple Silicon Mac, 8 GBBase MacBook Air M1 or M2, entry Mac mini | 30 s to 1 min 40 smeasured | 6 to 9 minmeasured | + 15 to 30 sestimated |
| Recent Windows laptop, 16 GB2023 or later, Core 12th gen or Ryzen 7000, 6 cores | 30 s to 1 minestimated | 3 to 6 minestimated | + 20 to 40 sestimated |
| 2019 Windows laptop, 8 GBCore i5 8th gen, 4 cores, one memory stick | 1 min 20 s to 1 min 45 smeasured | 6 to 10 minmeasured | + 30 s to 1 minestimated |
Detection loads about 1 GB of model into memory, Accurate mode another 1 GB. On an 8 GB machine with a browser, Teams and Outlook open, the operating system starts swapping to disk and the same page can take two to six times longer. We measured a single page swing from 2 seconds to over 2 minutes on the same MacBook Air for that reason alone.
Fast mode runs the detection model only and is what the figures above call detection. Accurate mode adds a small local language model that re-reads each passage, which costs roughly five to ten times more compute. The app times your machine at first launch and disables Accurate mode if it would be too slow to be useful.
A Word file or a PDF with selectable text is read directly. A scanned PDF or a photo is first rendered at 300 dpi and read by a local OCR engine, page by page, before detection starts. That step runs on the CPU on every platform, so it is where an older processor shows the most.
Sustained AI work heats the chip, and a laptop protects itself by slowing down. A fanless MacBook Air drops to roughly half speed after the first minute of Accurate mode. A thin Windows laptop on a soft surface does the same. Nothing is wrong, it is just physics, and it recovers as soon as the job ends.
None of these needs a setting inside the app. They are ordered by how much difference they make on an 8 GB laptop.
On battery, macOS Low Power Mode and the Windows power-saving plans cap the processor's clock speed. Plugging in removes that cap. Check the battery menu on a Mac and the power mode slider on Windows.
Before a large batch, close the browser tabs, video calls and Electron apps you do not need. On an 8 GB machine this is the single biggest speed-up, worth more than any other setting.
The first launch downloads about 1 GB of models, then runs a short calibration that times your hardware. On Windows that can take two minutes. Wait for the ready state before dropping your first file, and expect that first file to be a little slower than the next ones.
Loading the models takes 10 to 35 seconds each time the app starts. Leave it open during a working session and every file after the first starts instantly. A file you already processed in the same session comes back in about a second.
For most contracts, letters and spreadsheets, Fast mode finds the personal data you need to mask. Keep Accurate mode for documents where the wording around a name matters, and run those when you can leave the laptop alone for a few minutes.
If a document exists as a Word file or a text PDF, use that rather than a scanned copy. You skip OCR entirely, and the protected copy keeps its exact layout.
Put the laptop on a hard, flat surface rather than a lap or a bed, and keep the vents clear. On a fanless MacBook Air, a cool machine runs Accurate mode almost twice as fast as a warm one.
Some corporate antivirus tools scan every large file on read, which turns each model load into a minute-long wait. If the first launch or each start feels slow, ask IT to exclude the Hexagone AI application folder from on-access scanning.
The installer is small. The AI models are fetched once, on first launch, then everything runs offline.
Download of about 1 GB of models, so 1 to 5 minutes depending on your connection. Your Mac or PC receives the version built for its chip.
A calibration run that times detection and Accurate mode on your hardware. Up to 2 minutes on a 2019-class PC, a few seconds on a recent Mac.
The verdict is kept, so the next launches skip the calibration and are ready in 10 to 35 seconds.
The trial runs on the same models and the same hardware as the paid plans, so what you measure in the first week is what you get.