#Offline Translation
CatScribe is built local-first: documents are parsed, chunked, translated, reviewed, and exported on your machine, and the local engines send nothing to the internet. This page gives you the honest offline story — what runs fully offline, what needs an internet connection once, and how privacy is handled.
#The "Offline Translation" Option
In the default Simple interface mode, all local engines appear as one provider choice named "Offline Translation". Pick it, and the app runs the engine configured at Settings → Providers → "Preferred Offline Provider":
- Argos — the default and recommended engine; translates via downloadable language packs.
- LibreTranslate — a local server engine you run yourself in Docker.
- MarianMT and NLLB — neural models built into the app; NLLB covers 200+ languages.
When Offline Translation is selected, the selector shows "Currently using" with the resolved engine and an "Open Translation Settings" shortcut. To work with engines individually, switch to Advanced mode — see Simple vs Advanced Mode and Translation Engines.
#First-Run Downloads That Need Internet
"Offline" engines still need their models, and models are downloaded — plan one connected session before going offline:
| Download | Approximate size | When |
|---|---|---|
| Argos language packs | ~50–100 MB per language pair | First install of each pair, from the Argos Language Packs dialog or onboarding |
| MarianMT / NLLB weights | ~200–600 MB (NLLB model ~1.2 GB) | First use — "Models download on first use from Hugging Face … and stay cached locally." |
| COMET quality model (optional) | ~1–2 GB | Only if you enable COMET scoring in Settings |
| AI Subtitle Detection (optional) | ~250 MB | Only if you enable burned-in subtitle OCR |
| FFmpeg (subtitle/video features) | ~100 MB | Only if not already present |
| Ollama models (optional) | ~0.6–40+ GB depending on model | From the Providers/AI Models screen |
Two honest caveats:
- Out-of-the-box packs depend on the installer build. Some releases pre-seed the English→Portuguese Argos pack; a fresh install without a pre-seeded pack has no usable language pair until you download one. Check the Providers/AI Models screen or the Argos Language Packs dialog after installing.
- Downloads do not resume. If a pack or model download is interrupted, it restarts from the beginning — do large downloads on a stable connection.
After that, a machine can stay disconnected indefinitely: Argos and MarianMT/NLLB initialize and translate without any network calls.
#Setting Up For Fully Offline Work
- While connected, run the First Launch Setup and pick your engine ("Argos Translate" is the recommended card; "MarianMT / NLLB" is the built-in neural alternative).
- Install the Argos language packs for every pair you will need (each ~50–100 MB), or run one Marian/NLLB translation per pair so the weights get cached.
- If you want AI refinement offline, install Ollama (the app offers "Install Ollama Automatically") and download a model recommended for your hardware from the Providers/AI Models screen.
- Optionally install COMET (~1–2 GB) for quality scoring and any other optional dependencies you want.
- Disconnect and verify with a short test translation using Offline Translation.
#Performance Expectations
Offline translation uses your CPU, RAM, and disk. MarianMT/NLLB load models into system RAM (roughly 1–2 GB while active) and warn about high RAM usage on selection; Argos is the lightest starting point. If translation feels slow, use Auto Mode Fast, translate chapter by chapter, or see Performance Guidelines.
For the strategic case for offline pipelines, see the guide Offline AI Translation.