The questions, and the unflattering answers.

Yes. The documents, the search index and the language model are all files on the phone. Put the handset in airplane mode, pull the SIM, sit in a basement, and it behaves the same way it does on wifi, because it was never using the wifi.

The only things that need a connection are the first download and any later updates you choose to take. See how it works offline.

Slower than a cloud chatbot, and we are not going to print a number yet. A model thinking on phone hardware takes real time, and how long depends heavily on the handset.

We are actively working on it. When the figure is one we would stand behind on any phone we recommend, it goes on this page. Until then, treat anyone quoting you a speed for an on-device model as someone who has not measured it on your phone.

Yes, and so can any tool. Two things reduce the damage. First, the model answers from passages the search pulled out of the installed documents rather than from memory. Second, every answer names the document and page it used, and those citations open.

If something reads oddly, open the page and check it. We do not publish an accuracy percentage, because there is no honest single number for that and any site showing you one chose it.

It is a reference tool, not a medical provider. It can find what published guidance says and show you the page. It cannot examine anyone and it does not know your situation.

If emergency services or a clinician are reachable, that is the answer, every time. Use this for the case where help is genuinely not reachable, and read the source page rather than only the summary.

An ordinary Android phone, not a special device. It is running on regular handsets today.

Newer phones with more memory answer faster, and there is a floor below which the experience is not good. How much memory a handset has also decides which model sizes it is offered. We will publish specific minimum requirements and a tested handset list rather than a vague promise, once that list is finished.

On Android the default is Qwen3.5 at four billion parameters, because it is the size that fits the most phones. If your handset has the memory for it, an eight-billion or nine-billion build is offered as well: larger, more deliberate, and heavier on both memory and storage. Each size comes as the stock public checkpoint and as a variant tuned against this library.

The app looks at your device and suggests a size. You can take a different one instead, and it tells you what that will cost you rather than quietly removing the choice.

Both. Every model in the product is built on a stock, publicly available checkpoint, chosen and checked against this library, and each one is paired with a variant tuned on the library itself, judged blind against the stock build it started from. The two launch together.

What does not change between them is the rule. Every answer names the document and the page it used, because any model, tuned or stock, can write a confident sentence the library does not back, and a tuned build clears the same safety gates as the stock one before it ships. The full story is on what it runs on.

Gigabytes, and how many is up to you. The library is split by subject so you can install everything or just the parts that match your situation.

There is no way around the basic fact that a real library of real manuals plus a language model is a large amount of data, and we would rather say that here than surprise you at install time.

• 11.41GBEverything served
• 7.25GBThe documents
• 3.59GBThe maps
• 0.57GBThe search index

Measured off the reference build on 20 August 2026. The library page breaks the same total down pack by pack, with a running column so you can read down it until you reach the space you have.

No. There is nowhere for them to go. The model runs on your device, so the question text is processed locally and no outbound request is made to answer it.

That is a property of how the thing is built, not a promise in a policy document. The difference matters, because a policy can change and an architecture with no server in it cannot quietly start using one.

Those work, and people who run them are not wrong. The difference is where the thing is when you need it. A Pi, a mini PC or a drive is a second device you have to own, power, remember and reach.

The phone is already in your pocket and already charged. Getting a full library and a model to run properly on a handset is the part that has not been solved, and it is the part this is aimed at.

Yes, and that is where the largest version of it lives. On a desktop or laptop with a real graphics card or plenty of memory, SurvivalOps runs Qwen3.8-27B, a twenty-seven-billion-parameter model, over the same library, which is the base the whole project was built and checked against.

The desktop tier launches alongside the phone build rather than after it: the same documents and the same citations, with a much larger reader over them. The full picture is on what it runs on.

Not yet. The plan is the same shape as Android: Qwen3.5 at four billion or nine billion parameters, chosen by what the handset can hold, reading the identical library.

It is not built. There is no App Store listing, no build to test and no date, and we are not going to give you one we cannot keep. If you are on iPhone the waitlist is still the right place to be, and we will say plainly when there is something for you.

There is no price and no release date yet. Anyone telling you otherwise is guessing.

The waitlist is how you find out first, and joining it does not commit you to anything.

English today. The documents are English-language manuals and references, and the app expects questions in English.