Verta

Verta is a private, installable Arabic ↔ English translation workspace for documents and images. Translation runs on small open-weight models on your CPU, directly inside the browser. No API key, account, server inference, GPU, browser translation feature, or file upload is required.

Translation engine

Direction	Model	8-bit encoder + decoder
English → Arabic	Xenova/opus-mt-en-ar	110.9 MB
Arabic → English	Xenova/opus-mt-ar-en	110.9 MB

These are ONNX conversions of Helsinki-NLP's Apache-2.0-licensed English → Arabic and Arabic → English OPUS-MT models. Model revisions and Transformers.js 3.8.1 are pinned in app/local-translation/models.ts. Tokenizer/config files add a few MB per direction; the bundled WASM runtime is additional.

The 6 GB total system RAM target is supported by a conservative configuration:

• 8-bit model weights, CPU/WASM execution, and a single inference thread.
• One translation worker/model at a time, released after completion or cancel.
• OCR/PDF extraction finishes before the translation model loads.
• One sentence at a time, at most 192 source tokens including the language prefix. These models are trained on sentence pairs and can drop sentences when given a multi-sentence block, so Verta never sends one.
• Greedy decoding (num_beams: 1) and a 384-token output limit.
• An output-limit error is shown instead of silently accepting truncated text.

In CI (headless Chromium on Linux), the whole browser peaked at roughly 1 GB of proportional memory (PSS) while translating, sampled every 250 ms. That is a measurement on a CI machine, not a benchmark on a 6 GB device: available RAM, the operating system, other tabs, document size, and image OCR all matter. Close memory-heavy tabs and process large files separately.

These are sentence-level translation models, not general-purpose chat LLMs. They can miss idioms, terminology, and context across sentences, and they sometimes write small numbers as words. English → Arabic uses the >>ara<< target prefix. Preferred terminology uses best-effort text replacement, not constrained decoding or document-wide reasoning.

First use and offline use

1. Open Verta in a current browser with WebAssembly and Web Workers, over HTTPS or localhost. No browser translation feature or extension is needed. Optionally install it as an app from the browser menu.
2. On the first visit the service worker saves every file of the app, including the PDF reader, image OCR (English and Arabic) and Word exporter, so all features work offline later, not only the ones already used.
3. While online, press Save both · 222 MB to download both translation models from Hugging Face. (Translating in a direction also downloads that direction's model.) Only model files are fetched; your documents and translations never leave the device.
4. When the panel says both directions are saved, Verta translates Arabic ↔ English with no internet. Keep site data so the browser keeps the files.

Browser storage eviction, private browsing, or clearing site data can remove saved files. The status checks the actual cached model files and does not rely on a permanent “downloaded” flag. If a download was interrupted, connect and retry.

Windows app

Verta is also a Windows program. Download Verta-Setup-<version>.exe from the latest release and run it: it installs for your user account (no administrator rights) and adds Start menu and desktop shortcuts. Windows 10 or 11, 64-bit.

• Both models are inside the installer, so it translates with no internet from the first launch. Nothing is downloaded after installation.
• The app serves its own files from 127.0.0.1 inside an Electron window and blocks every request to another host, so documents never leave the PC.
• The installer is not code-signed, so Windows SmartScreen may show "Windows protected your PC". Choose More info → Run anyway.

The Windows app workflow (.github/workflows/desktop.yml) builds the installer on Windows, downloads the pinned models and checks every file against Hugging Face's published hashes (scripts/fetch-models.mjs), installs the app silently, and runs tests/desktop.e2e.mjs against the installed Verta.exe. Pushing a tag such as v1.0.0 (matching desktop/package.json) publishes the installer as a GitHub Release. The Electron shell is in desktop/main.mjs.

Supported files

• PDF, DOCX, PPTX, ODT, RTF, plain text, Markdown, and CSV
• PNG, JPG, JPEG, and WebP images with English or Arabic source text. For photos, lines the OCR reader is unsure of (logos, frames, walls, blur) are skipped so they are not translated as nonsense, wrapped sentences are rejoined, and ALL-CAPS English is translated in sentence case. The result says how many unclear lines were skipped.
• Word (.docx) and plain-text output

Legacy binary .doc and .ppt files are not supported in offline mode. Document sections are reconstructed as text; complex source-page layout is not preserved. Scanned PDFs need OCR outside the PDF text-extraction path.

Run locally

Requirements: Node.js 22.13+ and pnpm 11+.

pnpm install --frozen-lockfile
pnpm dev

Open the local URL printed by the server. Build scripts bundle the translation worker and copy its WASM runtime, PDF worker, and OCR assets into public/offline. Weights are downloaded on demand, not bundled into the repository or site build. The optional native onnxruntime-node install script is disabled because the application uses the browser's WASM backend.

Checks

pnpm run typecheck
pnpm run lint
pnpm test
NEXT_PUBLIC_STATIC_EXPORT=true pnpm run build:github

The Verify Verta GitHub Actions workflow runs these on every push, plus pnpm run test:browser: the real pinned models in Chromium with the browser's own Translator API removed. It covers both directions online, an offline reload, the Save both button followed by a first-ever Arabic → English job offline, long numbered input that fails if any sentence is missing, a .docx upload and Word export while offline, and cancel-then-restart.

The unit suite covers sentence and token-budget splitting without source loss, output-limit handling, worker cancellation and cleanup, one active model, cache eviction, engine routing, same-language extraction, and offline error handling. Unit worker doubles exercise the application contract; they do not measure model quality. See model validation for real-model smoke tests and their limits.

GitHub Pages

The included workflow publishes the static app whenever main is pushed. GitHub Pages must use GitHub Actions as its publishing source. The worker and WASM assets respect the repository subpath.