What it runs on
One library. Three sizes of machine.
The documents are the same on every device. The index is the same, and so are the citations. The only thing that changes from a desktop to a phone is how large a model is doing the reading, and that is chosen to fit the hardware you have. This page names those models, and the tuned variants that launch beside them.
01Scales
The same product at three sizes
Not three products that happen to share a name.
-
Full strength
Desktop and laptop
Qwen3.8-27B, twenty-seven billion parameters, and the base the whole project was built and checked against. It wants a real graphics card or a generous amount of memory, and in exchange it can hold more of the library in view at once.
At launch it installs on a desktop or laptop the same way the phone build installs on a handset. The same documents, the same index, a much larger reader over them.
-
The lead tier
Android phones
Qwen3 and Qwen3.5. The default is Qwen3.5 at four billion parameters, because it is the build that fits the most handsets in ordinary pockets. Phones with the memory to spare are offered an eight-billion or nine-billion build instead, larger and more deliberate.
The app looks at the device, suggests the size that suits it, and lets you take a different one instead. It says what that will cost you rather than hiding the choice.
-
Planned
iPhone and iPad
Qwen3.5, at four billion or nine billion parameters, picked by what the hardware can hold. Apple devices vary enough that the model has to be chosen per device rather than per platform.
This one is not built. There is no App Store listing, no build to test and no date. We would rather leave the row empty than fill it with a promise.
Every size above ships twice, once as the stock public checkpoint and once as a variant tuned on this library. The planned tier is drawn empty because there is nothing to measure yet.
This is a drawing of size, and size only. A parameter count says how much machine a model needs, not how right it is, and this page does not rank the tiers against each other. What every bar on it shares is the library underneath: identical documents, identical index, identical citations.
02Spec
The launch configuration
Written down as a spec, so nobody has to infer it from a screenshot.- Documents and index
- Identical on every tier
- Citations
- Resolve locally on every tier
- Network required
- None, on every tier
- Desktop and laptop
- Qwen3.8-27B
- Android models
- Qwen3 and Qwen3.5
- Android default
- Qwen3.5 at 4B
- Android options
- 8B and 9B, memory allowing
- Model variants
- Stock and survival-tuned
- iPhone and iPad
- Qwen3.5 4B or 9B, planned
The three rows at the top are the product. The rows below them are an implementation detail that we happen to think you deserve to see.
03Tuning
Stock models, and the tuned variants beside them
Two readers per tier. One rule for both.Every model named above is a stock, publicly available checkpoint, chosen and checked against this library. None of them launches alone. Each is paired with a tuned variant of itself: the same model, trained further against this exact library, and measured against the stock build it started from in blind comparisons scored by independent judges.
The tuning exists because a general model and a survival library start out as strangers. A variant trained on these documents learns their shape: where the tested numbers live, which procedures allow no paraphrase, and when the only right answer is the page itself. That is what the blind rounds measure, and it is why the tuned builds belong in the product rather than in a lab drawer.
One rule holds for both kinds of reader, and it is the oldest rule in the product. Every answer on every tier names the document and the page it came from, and the citation opens, because any model, stock or tuned, large or small, can write a confident sentence the library does not back. Tuning does not retire that rule and nothing will. A tuned variant clears the same safety gates as the stock model before it ships, and it answers under the same citation discipline after, so the page it names is always there to check it against.
04Rationale
Why the model is the only part that changes
Because it is the only part that has to.A library is files. An index over that library is a file too. Neither of them cares whether it is sitting on a desktop with a graphics card in it or on a handset in a coat pocket, so neither of them is cut down to fit a phone. What you install on the phone is the library, not a summary of the library.
The model is the part that has to fit, because it is the part that has to run. A larger model has more room to be careful with a long or awkward question, and a smaller one writes a shorter and plainer answer. That is the whole reason there is more than one size, and it is why the citation matters more on the small tier than on the large one, not less.
Which is the point of the citation in the first place. On every tier, the answer names the document and the page, and on every tier that page is a file already on your device. If the model was terse, or you think it missed something, the source is one tap away and it does not change with the model that pointed you at it.
- 1,421Documents
- 122,509Indexed passages
- 80,155Pages with indexed text
- 15Subject packs
Not the phone's share of the library. The whole of it. That is the same count of documents and the same count of passages a desktop reads, because the files that install on a handset are the files, and the index over them is the same index.
The same library and the same answers everywhere, with the model sized to the device. A question you ask on the laptop and the same question on the phone are being answered out of the same pages. The larger machine reads them with more room to think. It is not reading a different book.
05Disclosure
What this page is not telling you
The comparisons we could draw and will not.- No speed comparison between the tiers. The large model on a desktop is faster than the small model on a phone, and everyone reading this already worked that out. Putting figures on it would mean publishing phone response times, and those are not numbers we are willing to stand behind yet.
- No accuracy ranking between the tiers. We are not going to hand you a table saying one size is some percentage better than another. There is no honest single number for that, and the answer you should be checking is the citation anyway.
- No date. Desktop and Android are the launch tiers, with stock and tuned models on both. iPhone and iPad come after, and we will not pretend to know the calendar better than we do. When a date exists, the waitlist hears it first.
- No score for the tuned variants. The blind comparisons are real and we could print the numbers, but a bare score with no rubric attached is exactly the kind of figure this site refuses to publish. What we will publish at launch is how the tuned builds were gated, not just that they won.
- No minimum specification yet. There is a floor on every tier below which this is not a good experience, and we would rather publish a tested list of what works than a hopeful number that leaves you out of pocket.
06Continue