iPhone Apps Whose AI Runs On-Device

Most apps advertising AI are sending your input to somebody else's server. That's not automatically wrong, but it is a materially different product: it needs a connection, it has a per-request cost that someone eventually pays, and your text or photos leave the device. Three of our apps run their AI entirely on the iPhone instead. This page explains what that means technically, which devices support it, and exactly what happens on the ones that don't.

On-device vs cloud, in practical terms

Apple ships two things worth distinguishing. Apple Foundation Models is the on-device language model that apps can call directly through the Foundation Models framework — it runs on the phone's own silicon, with no request leaving the device. Vision is Apple's on-device image analysis framework, which does things like subject masking locally. Neither involves a server round trip.

The consequences, in order of how much they'll affect you:

• Your input stays put. Not "encrypted in transit", not "deleted after 30 days" — it doesn't leave. There is no Solutionflare server for it to reach, so there is no retention policy to trust and no breach that could expose it.
• It works with no connection. A cloud AI feature is a network feature. An on-device one keeps working in airplane mode, which is why our AI features survive the airplane-mode test.
• There's no per-use meter. Server inference costs money per request, which is why cloud-AI apps tend toward subscriptions, credit packs, and daily limits. On-device inference costs us nothing per use, which is why Dogear can generate unlimited passages and quizzes and still be entirely free.
• It's genuinely more limited. On-device models are smaller than frontier cloud models. For the jobs we use them for — writing a short passage, generating four comprehension questions from a text you just read, summarising a week of sleep — that's the right size. We wouldn't claim it for everything.

Which of our apps use AI, and for what

Dogear — passages and comprehension quizzes

Dogear is a speed-reading trainer whose whole premise depends on this. After every passage it generates a short multiple-choice quiz from that exact text, scores it, and explains each answer — which is how it reports your comprehension-safe speed, the fastest pace at which you still scored 80% or better. It can also generate fresh passages to read.

That means the AI is reading whatever you read, including anything you paste in from work. Running it on-device is the only version of this feature we were willing to ship. Dogear is free, has no account, and its App Store privacy label is "Data Not Collected".

Without Apple Intelligence: Dogear ships bundled passages that each include a real, hand-built comprehension quiz, so the pace-and-prove loop works on every supported iPhone. You can also paste and read any text at any time. Generation of new passages and quizzes is the part that needs the model.

Rouse — sleep insights and morning briefings

Rouse analyses your sleep patterns to produce nightly insights, weekly through all-time reports, and a morning Rise Briefing. Sleep and health data is about as sensitive as phone data gets, and it is read and written only under HealthKit permissions you grant and can revoke in Apple's Health app.

All of that analysis runs on-device through Apple Foundation Models. There is no cloud AI in Rouse, and no sleep or health data leaves the phone — including to us.

Without Apple Intelligence: Rouse produces built-in template summaries instead, also generated on-device. You still get reports; they're written from templates rather than composed.

Pentire — turning a photo into a puzzle

Pentire cuts any photo into numbered sticker pieces using Apple's on-device Vision subject masking plus its own segmentation, merge, and contour pipeline. This is image understanding rather than a language model, and it is the feature people most expect to involve an upload — every photo-to-paint-by-number website on the internet works by sending your picture to a server.

Pentire doesn't. There is no image-processing server, and photo import works in airplane mode, which is the practical proof rather than a promise. Your photos are never uploaded, which also means photo puzzles are the one thing that can't be shared by code — the code is a recipe, and the recipe would need the picture.

Which iPhones can run it

Apple Intelligence requires an iPhone 15 Pro or newer. Both Rouse and Dogear also require iOS 26 or later, which is where the Foundation Models framework and AlarmKit live.

On an older but supported device, neither app breaks — each has a designed fallback, listed above. This matters more than it sounds: an app that treats on-device AI as required rather than enhanced simply refuses to work for most of the installed base, and an app that silently falls back to a cloud model has quietly undone the privacy property you installed it for. We do neither.

Pentire's photo cutting uses Vision rather than Apple Intelligence, so it runs on the full iOS 17+ range with no special hardware requirement.

What we don't do with AI

• No cloud inference on user content. No text, photo, sleep record, or message from any of our apps is sent to an external model provider.
• No AI-as-a-price-lever. Nothing generates a credit meter, a daily quota, or a "get more generations" purchase. Dogear's generation is unlimited and free.
• No AI where it isn't better. Chat Local, CalcBoard, and Scroll Clash contain no AI features at all, because none of those problems is improved by one. A calculator should be exact, not creative.
• No training on your data. We receive none of it, so there is nothing to train on. See the complete list of what each app sends.

Frequently asked questions

Which iPhone apps run AI entirely on the device?

Three Solutionflare apps do. Dogear generates reading passages and comprehension quizzes with Apple Foundation Models. Rouse generates sleep insights, reports, and morning briefings the same way. Pentire cuts photos into sticker puzzles using Apple's on-device Vision subject masking plus its own segmentation pipeline. None of them sends user content to a cloud model.

What is the difference between on-device AI and cloud AI?

On-device AI runs the model on the phone's own silicon, so your input never leaves the device, the feature works with no connection, and there is no per-request cost. Cloud AI sends your input to a server, which requires a connection, creates a data-retention question, and costs the developer money per use — which is usually why such apps have subscriptions, credit packs, or daily limits. On-device models are smaller and less capable, so the trade-off is real in both directions.

Which iPhones support Apple Intelligence features in these apps?

Apple Intelligence requires an iPhone 15 Pro or newer, and both Rouse and Dogear also require iOS 26 or later. On devices without Apple Intelligence, Dogear uses bundled passages that each ship with a real hand-built quiz, and Rouse produces built-in template summaries, also generated on-device. Pentire's photo cutting uses the Vision framework rather than Apple Intelligence, so it works on any supported device from iOS 17.

Does Dogear send what I read to a server?

No. Passages and quizzes are generated by Apple's on-device foundation model, so the text you read — including anything you paste in — never leaves your iPhone. There is no Dogear server. Its App Store privacy label is "Data Not Collected", and the app requires no account.

Does Rouse send my sleep or health data anywhere?

No. Every AI feature in Rouse — nightly insights, weekly through all-time reports, and morning briefings — runs on-device through Apple Foundation Models. Health data is read and written only under the HealthKit permissions you grant, which you can revoke at any time in Apple's Health app. The only network requests Rouse makes at all are a static settings file at launch, Apple's WeatherKit for weather, and App Store purchase processing.

Are my photos uploaded when Pentire makes a puzzle?

No. Cutting runs entirely on the device using Apple's on-device Vision subject masking plus Pentire's own segmentation and contour pipeline. There is no image-processing server, and photo import works in airplane mode, which is the practical way to verify it.

Do the AI features cost extra or have usage limits?

No credits, quotas, or per-generation charges exist. Dogear's generation is unlimited and the whole app is free. Rouse's on-device sleep insights are part of Premium, which is a flat $2.99 a month, $24.99 a year, or $59.99 once — not a metered AI charge. Pentire's photo cutting is limited by tier (two photo puzzles free, unlimited with Plus), not by generation count.

Do any Solutionflare apps use ChatGPT or another cloud model?

No shipped feature in any of our apps sends user content to an external model provider. AI features run through Apple's on-device frameworks, and three of the six apps — Chat Local, CalcBoard, and Scroll Clash — contain no AI features at all.