WitNote 2.0.1 remains the existing App Store product record, but its implementation has gone through a substantial native Swift rewrite. The goal was not to add noise; it was to bring Markdown writing, local files, format conversion, bilingual reading, and on-device AI back into one inspectable workflow.

The release gate therefore covers both engineering and store truth: a clean Mac must be able to start local AI, folder authorization must recover after a restart, model downloads must cancel and reset cleanly, Lifetime purchases must restore, and third-party model notices must ship with the package.

A native macOS workbench

The workbench, folder access, editor, preview, chat, model management, format conversion, and settings now have clearer Swift-native boundaries. Folder authorization must account for security-scoped bookmarks, restart recovery, and an explicit re-authorize path when the saved permission is no longer usable.

This makes WitNote a desktop workbench rather than a web surface in a window. Window lifecycle, menus, file permissions, cancellation, and error feedback are part of the product experience.

Local Markdown remains the source of truth

Markdown files stay at the center. The folder and document list organize work, the editor changes it, preview and split view support reading, and conversion can deliver PDF, Word, EPUB, or another workflow. EPUB and bilingual reading are supporting capabilities; the product is positioned as a local Markdown writing and knowledge workspace.

The content path stays traceable: the user chooses the files and directory, the app works inside that permission boundary, and import, conversion, and export should not take control away from the original files.

Local MLX and model management

The AI route uses local MLX inference. A lightweight Qwen3 0.6B 4-bit starter model is bundled so a new user can complete a local AI action before downloading anything else. Additional models are optional downloads, shown with hardware guidance and real file status.

The downloader separates received bytes, speed, remaining time, network waiting, cancellation, failure cleanup, and a fresh retry. An interrupted directory must not be mistaken for an installed model. Model selection refreshes across chat, bilingual translation, and other AI entry points without requiring a restart.

Privacy, purchases, and third-party boundaries

Local files and local inference are the default boundary. Model downloads, Apple StoreKit, legal and support pages, and external links opened by the user are distinct network actions. Apple handles the Lifetime purchase and restore flow; WitNote only uses the StoreKit result and does not receive Apple ID or payment details.

MLX, Qwen, and subsequently downloaded models are not WitNote-owned assets. The package provides third-party notices and bundled model license files, while the website and About surface identify model sources without overstating their ownership or capabilities.

The release gate

The 2.0.1 gate includes 1,300/1,300 automated tests, Distribution signing and a fixed SHA-256, a no-tools and no-cache first launch, Lifetime IAP purchase and restore, twelve-language checks, privacy and license notices, cancellation, memory, and long-text behavior.

These checks keep the website, App Store metadata, and submitted package aligned. A local AI feature that cannot be reproduced in a clean environment is not presented as delivered, and unverified performance is not turned into an absolute promise.