Does an on-device cigar scanner actually keep my photos and data private?
It depends on where the recognition happens, not on what the app calls itself. If the model that identifies your cigar runs on your phone, the photo is processed and discarded locally, never sent anywhere. If the app posts the photo to a server instead, it has left your device, and you can usually tell which one you're using.
What does "on-device" actually mean, technically?
A cigar-identification model has to run somewhere. It reads the band text, the wrapper colour, the shape of the cap, and the proportions of the cigar, and turns that into a match. There are only two places that computation can happen: on the hardware already in your hand, or on a remote server that receives an image you've sent it.
On-device means the first one. The model itself, not just a link to your photo, is shipped inside the app and executes locally. In a browser, that looks like WebAssembly running a model in the ONNX format. On iOS, it's Apple's own CoreML and Vision frameworks doing the same job with native code. Either way, the pixels never need to cross a network to produce an identification.
Paradigm's scanner works this way: matching runs as WebAssembly and ONNX in the browser, or through CoreML and Vision on iOS, and the identification step doesn't require a server call at all.
How can I tell if a scanner is uploading my photo or not?
You don't have to take a privacy policy's word for it. A few checks you can run yourself:
- Turn on airplane mode and try to scan. An on-device app keeps working, because the model is already sitting on your phone. A cloud-based one will hang, error out, or fall back to asking you to type the name in manually.
- Watch what happens on first use. On-device apps typically have a noticeable one-time download or load, because the model file itself has to arrive before it can run locally. Apps with no such step, that stay small and always need a live connection to identify anything, are very likely sending the image out.
- If it's a web app, open the browser's network tab while you scan. A request that fires with an image attached as it's identifying is the photo leaving your device. No such request means the work happened locally.
- Read what the app says about itself, literally. "AI-powered" tells you nothing about where the AI runs. "Runs on your device" or "no cloud processing" is a specific, checkable claim; vague enthusiasm about accuracy usually isn't.
Does "on-device" mean the app never sees any of my data?
No, and it isn't meant to. On-device recognition is a claim about one specific step: identifying what's in the photo. Once you've identified a cigar and choose to log a rating, add it to your collection, or write a tasting note, that's a deliberate save, and it has to be stored somewhere for you to search it later, sync it across sessions, or see it again next week. That's not a privacy problem by itself, because you asked for it.
The distinction that matters is between data you chose to keep and data the app moved without telling you. A photo quietly uploaded for identification, with no record of that step in the interface, is different from a note you typed and hit save on. The strongest version of on-device processing goes further than just running locally: it extracts the identification result and discards the image afterwards, rather than keeping a copy anywhere once the match is made.
On-device vs cloud-upload: what actually differs?
| On-device recognition | Cloud-upload recognition | |
|---|---|---|
| Where the photo is processed | Your phone's own CPU or GPU | A remote server |
| What leaves your device | Nothing, unless you save a result yourself | The photo itself |
| Works with no signal | Yes, once the model has loaded once | No |
| How you verify it | Airplane-mode test, network tab | You're trusting the vendor's word |
| Model size | Limited by what a phone can hold and run | Effectively unlimited |
Is on-device recognition as accurate as sending the photo to a server?
Model size is a genuine tradeoff. A model that has to fit inside a phone's memory and battery budget is smaller than one running unconstrained on a server, and smaller models can miss things larger ones catch. That's real, and it's worth understanding in more detail: see how accurate AI cigar recognition actually is for what typically goes wrong and why.
But raw model size isn't the only, or even the main, variable. What the model is looking at matters more than where it runs. A scanner that only reads the printed band fails outright the moment that band is torn, soaked, or missing, regardless of how big the model behind it is. One that reads the wrapper, shape, and proportions alongside the band has other signals to fall back on. That failure pattern is worth its own look: see why band-only identification breaks down.
Matching what the model reads against a catalogue is also just a local lookup, not a reason to phone home on its own. Paradigm's catalogue currently holds 11,519 cigars; checking a read against a list of that size is database work, not something that needs cloud compute.
What should I actually check before trusting a scanner with cigar photos?
Five minutes, not a legal review:
- Skim the privacy policy for one sentence: does it say photos are uploaded anywhere, and if so, for what and for how long.
- Run the airplane-mode test once, on a cigar you don't mind re-scanning if it fails.
- Notice whether the app asks for camera permission every time you scan, or holds a background connection open beyond that.
- Check whether the same claim shows up more than once, in the store listing and inside the app itself, or whether it only appears in marketing copy and nowhere in the actual product.
Scanning privacy is one piece of a bigger decision. If you're weighing which app to trust with an ongoing record of what you've smoked, the fuller checklist is in what to actually look for in a cigar tracking app.
The short version
- On-device means the identification model runs on your phone; nothing has to leave it to name your cigar.
- A server-based scanner uploads your photo somewhere else to be matched, whatever the app calls that step.
- Test it yourself: airplane mode, then try to scan. On-device apps keep working; cloud-based ones don't.
- Saving a note or rating you typed is different from an app silently uploading a photo. One is a choice you made; the other isn't.
- Model size is a real accuracy tradeoff either way, but what the model reads (band alone versus the whole cigar) matters more than where the computation happens.