Understand any PDF without reading all of it.

A free tool that summarizes documents, builds outlines and entity lists, and answers questions about them with a language model that runs on your device — no API keys, no uploads, no per-page fees. It is aimed at students facing a reading list, researchers triaging papers, and professionals extracting the facts from long reports and contracts.

PDF, DOCX, TXT, and Markdown all work, two small open-weight models are on offer, and every run is saved in a local history you can return to weeks later.

How it works

The document’s text is extracted on-device — pdf.js walks a PDF page by page, DOCX is unpacked locally, and text and Markdown are read directly. Documents that fit one context window are analyzed in a single pass; longer ones are split at page boundaries into about 8,000-character parts. Each part is summarized separately (the map step), and the part summaries are distilled into one final brief (the reduce step), so the output covers the whole document rather than just its opening.

Answers stream token by token as the model generates, with live progress for the part-by-part phase. Everything is exported as ready-to-paste Markdown.

Six tasks, two models

Six task modes cover the common jobs: Summary starts with a one-paragraph overview followed by the most important points; Key points pulls facts, figures, dates, and action items into a bullet list; Outline produces a hierarchical Markdown outline of the full structure; Entities groups people, organizations, locations, dates, and terminology; Explain simply rewrites the document in plain language a high-school student can follow; and Ask a question answers from the document’s own content, declining when the answer is not in it.

Two quantized (q4f16) ONNX models run locally through Transformers.js: HuggingFaceTB/SmolLM2-1.7B-Instruct (about 1 GB) and onnx-community/Qwen2.5-1.5B-Instruct (about 1.1 GB). Both answer in the document’s language unless your question is phrased in another one.

Local run history

Every successful run — the document name, task, model, question, elapsed time, page count, and full output — is saved to your browser’s local storage. The history panel lets you restore an earlier brief, copy from it again, or delete individual entries, and nothing is synced to an account because there is no account. The free tier keeps the 20 most recent runs; Pro keeps 100, with the oldest entry rolled off automatically.

WebGPU, CPU mode, and scanned PDFs

Chrome or Edge with WebGPU is strongly recommended: generation is fast enough to feel conversational. Without WebGPU the models run on CPU via WebAssembly — every task still works, but a long document can take a good while, and the interface warns you up front rather than pretending otherwise. Scanned, image-only PDFs have no text layer to read; the tool detects this and directs you to the OCR tool first. Models download once from the open model hub and are cached for offline use.

Free and Pro limits

The free tier processes documents up to 120 pages with unlimited runs and a 20-item local history. Pro raises documents to 600 pages and history to 100 runs. A document over your limit contributes its first pages with a clear notice of how many were skipped — there is no hard file upload because nothing is uploaded at all; the practical ceiling is your device’s memory and patience. Both tiers include commercial use, carry no watermark, and have no monthly quota.

Honest expectations and privacy

A 1.5–1.7 billion-parameter local model summarizes faithfully and answers questions grounded in the text, but it is not a frontier cloud chatbot: treat its output as a fast first pass for open-ended reasoning and verify numbers in the source. Every model card — SmolLM2 and Qwen2.5 — is public on the model hub.

Papers, contracts, medical documents, and internal notes never leave your machine after the one-time model download. That is the core point of the tool: material you would never paste into a cloud LLM can be summarized safely.