Read deeply.
The chrome disappears, and what is left is the page. Fluid line length, generous leading, a typographic title bar. The PDF has the room it deserves; everything else is a whisper at the edge.
Quire is a research workspace for PDFs. A native reader where your highlights, your notes, and the AI you brought yourself all live in the margin of the page — attached to the sentence that spoke first. Not a chat panel. A footnote.
Recurrent models typically factor computation along the symbol positions of the input and output sequences. Aligning the positions to steps in computation time, they generate a sequence of hidden states, as a function of the previous hidden state and the input for position t.
This inherently sequential nature precludes parallelization within training examples, which becomes critical at longer sequence lengths.
Memory constraints also couple example batches to sequence length, and become more severe for longer sequences. Fundamentally, this has prevented the gains in compute efficiency that more parallel architectures enjoy — a bottleneck this paper proposes to remove entirely.
Every reader of every important book has, at one time or another, written in its margins. The marginalia rail is Quire's spine: a thin column to the right of the page where your highlights, your notes, and the model's responses all sit attached to the sentence that produced them.
Hover a footnote, the source pulses in the page. Tomorrow, when you reopen the document, every mark and every answer is exactly where you left it — written into a JSON sidecar that lives next to the PDF, in plain text you can read with your eyes.
The chrome disappears, and what is left is the page. Fluid line length, generous leading, a typographic title bar. The PDF has the room it deserves; everything else is a whisper at the edge.
With the keyboard or the trackpad — four ink colors, notes anchored to the words themselves. Annotations live beside the file in plain JSON: portable forever, conflict-friendly, yours.
In the margin of the page, not in a chat panel. Select a sentence, press ⌘E, and the response streams into a footnote attached to where you asked from. Selection-scoped by default; your key, your model, your bill.
The first day a tool works. The second day, the small decisions are what keep you. These are some of the second-day decisions.
OpenAI, Anthropic, Google, or Ollama. Your keys live in the system credential store — never on disk in plaintext, never on our servers (we do not have servers).
Highlights and notes are stored beside the PDF as .quare.json — plain text, diffable, git-friendly, copy-with-the-file portable.
Standard menu bar, platform-native shortcuts, settings open where they should. Quire feels like a citizen of the OS it runs on, not a wrapped website.
⌘ ,The default ⌘E request sends your selection plus about five hundred characters of surrounding context. Whole-page and whole-document modes are opt-in and warned.
⌘ EToggle "Use local model" and Quire routes through your Ollama install. Nothing leaves your machine. Reading on a plane stays reading on a plane.
One window per document. Restored on relaunch with selection, scroll position, and current marginalia view. Because that is how scholars actually work.
⌘ NPreview from your OS file manager with annotations baked in. The highlight on page four shows up in the thumbnail.
Open, jump, ask, switch provider — every action is one ⌘K away. No tab bar, no second toolbar to learn.
⌘ KA researcher, a graduate student, an analyst, a curious reader. Someone who sits at a desk with a lamp and a stack of papers — and fourteen tabs of PDFs open in a browser that should not be where they read.
Quire is not a Notion clone, not a Zotero replacement, not a wrapper around a chat model. It is a place to read carefully, mark what matters, ask the document to explain itself, and come back tomorrow to find the thread exactly where you dropped it.
We are admitting readers a few at a time, in order of arrival, and writing to each one personally. Tell us how you read, and we will tell you when it is your turn.
macOS first · Windows and Linux after the v1 settles.