Four ways Ask runs. None of them is a cloud.

Ask answers plain questions about your Apple Health data. However it runs, your iPhone reads the data and computes every number. A language model, when one is used, only writes the sentence, and a sentence with any number your iPhone didn't compute is thrown away.

Where your data goes

Your iPhone, every mode

1. ReadApple Health, read-only. Nothing is written back.
2. UnderstandWhich metric, which days, which calculation.
3. ComputeEvery number, by the same code in every mode.
4. Word and checkTemplate or model. Any unmatched number means the template is used.

Your Mac, optional

One request per answer: the question and that answer's few numbers, over your Wi-Fi. Back comes one sentence, checked on the iPhone.

Not released yet.

✕ No cloud AI, no API keys Never sent anywhere: your health history, the app's data cache, or anything to us.

The four modes

You don't pick from a menu. Your iPhone uses Apple's model where it can. Where it can't, Guided mode answers, and you can add a downloaded model, or your own Mac once that ships. The engine that answered is named under every answer.

• 1Guided mode In 1.19

  Shown as "Answered on your iPhone · Guided mode"

  No model at all. A parser on your iPhone reads the question, the app computes the answer, and a fixed template states it. The arithmetic is the same code the other modes use, so the numbers are the same. If it misreads you, tap a chip to correct the metric or the dates.

  It isn't AI, and the app never calls it AI.

  What leaves your iPhone: nothing. Works in Airplane Mode.

• 2Apple's on-device model In 1.19

  Shown as "Answered on your iPhone · Apple Intelligence"

  Apple's built-in language model reads the question, picks the calculation and words the result. The numbers still come from the app's own code. If the model writes a figure the app didn't compute, its sentence is discarded and the template is shown instead.

  The app uses Apple's on-device model only, not Private Cloud Compute.

  What leaves your iPhone: nothing. Works in Airplane Mode.

• 3A downloaded model In 1.19

  Shown as "Answered on your iPhone · On-device model"

  For iPhones without Apple Intelligence. An optional small model (Qwen2.5 0.5B, fine-tuned for this one job) that you download once in Settings. Guided mode still reads the question and computes the answer; the model only rewords it, behind the same number check.

  Remove it any time to free the space.

  What leaves your iPhone: nothing. The model file comes down once from our model host, over Wi-Fi only; that request carries nothing about you or your data.

• 4Your Mac Not released

  Will show as "Numbers computed on your iPhone · Worded by your Mac"

  For iPhones without Apple Intelligence that live near a Mac. Ollama on your own Mac rewords the answer, with a bigger model than fits on a phone if you like. Your iPhone still reads the question and computes every number, and the same check throws out any sentence with a number it didn't compute.

  A consent screen explains all of this before anything is sent, and Settings › What was sent shows the last request, byte for byte.

  What leaves your iPhone: the question and that one answer's numbers, to the Mac you name, over your Wi-Fi (or your own Tailscale network). Never your health history, and never to us.

Side by side

Ask, the downloaded model and Your Mac are part of the one-time unlock. "Checked" means a sentence with any number the iPhone didn't compute, or one that leaves out the answer, is discarded and the template is shown.

Why the model can't change a number

Small language models are fluent and not reliable with figures. So none of these models is trusted with arithmetic. Every answer is computed first, then the model gets only that answer's facts: the metric, the date range, the result and the day count. Before its sentence is shown, the app checks three things:

• every number in the sentence is one the app computed,
• the sentence actually states the answer, and
• it describes the direction correctly (up is up).

If any check fails, you see the template instead, and the label under the answer says which engine wrote it. Tap the answer to see the days and samples it was built from.

Setting up Your Mac, when it ships

• On the Mac: install Ollama, pull a model (for example ollama pull llama3.2), and turn on Expose Ollama to the network. Ollama listens only on the Mac itself until you do (Ollama FAQ).
• On the iPhone: type the Mac's local name, such as studio.local, or its name on your own Tailscale network. Raw IP addresses are refused, because iOS blocks plain connections to them and they change.
• Pick a local model. Ollama can also run cloud models. A local one keeps the request on your Mac.
• Use a network you trust. Ollama has no password, so anyone on the same Wi-Fi could use it too. The app says so before you turn it on.

Common questions