Languages, model downloads & offline availability
The first local MVP includes English and Spanish, using pinned Tesseract.js 7 and 4.0.0_best_int language packages.
Explicit downloads
The English language pack is about 2.8 MiB and Spanish about 2.0 MiB. Initial loading also downloads roughly 11.3 MiB of shared engine/worker assets covering baseline, SIMD and relaxed-SIMD WASM variants. Actual bundled byte counts are shown in the studio. Byte progress includes reused cached assets; integrity is verified before OCR becomes ready. All assets are self-hosted.
What offline means here
On supported HTTPS/localhost browsers, Load language assets registers an asset-only service worker and stores only pinned OCR resources in Cache Storage. The app checks byte size and SHA-256. Assets available offline means OCR can run in this already open page without the network. The app shell is not cached, so reopening or reloading offline is not supported. Browser storage eviction or private mode can remove/prevent the cache.
Remove or retry
Clear workspace removes image/text references but keeps reusable OCR assets. Remove offline assets deletes this build’s OCR cache and unregisters this app’s cache worker. Missing/corrupt assets show a retry path: remove assets, reconnect if necessary, reload the selected language assets and re-extract. No remote OCR fallback exists.
Quality scope
The documented synthetic printed-text corpus contains three clean fixtures per language. Its normalized character-error-rate gate is at most 5% per fixture. This does not establish real-photo, handwriting, multi-column, mobile-device or general-language accuracy. Unsupported languages are not selectable; if the source is another language, use a tool with a reviewed model for it.