Private machine translation you can run yourself

LibreTranslate is an HTTP API another program can call. The README says it is a self-hosted API and that the engine is Argos Translate, not Google or Azure. “Offline capable” is the project’s phrase. You still run the process: Python 3.8 or newer, or Docker.

A default install loads every available language. If a direct pair is missing, it pivots, usually through English. The FAQ adds a caveat in the next breath: the public server at libretranslate.com sometimes runs models that are not in the Argos index yet, including some converted from OPUS. A copy you install can lag that server.

The public API is not free. The docs say a key is required there, and that your own server is the free path. The portal lists Pro at $29 a month, about 20 continuous requests a minute and 2,000 characters each, and Business at $58 a month, about 50 continuous requests a minute, with the same character cap. The burst numbers on that page are higher, 80 and 200, and the page says those are bursts. We’d recommend LibreTranslate if the caller needs that HTTP API on a machine you operate and the text has to stay on that machine. An in-process Python call is Argos Translate, not this server.

TranslateLocally is the window. The README says translation happens on your local machine, using Marian and Bergamot. The site translatelocally.com redirects to private.mt, which calls it a privacy-focused desktop app and says it is free, then says extra functionality and language support are by contact. No price is printed for those extras.

Builds are published for Windows, Linux, and Mac, with compatibility builds for old hardware and M1. Qt, libarchive, and Intel MKL show up if you compile. The README discourages OpenBLAS and says it is much slower than Intel MKL.

Other programs can talk to it through native messaging, which is not the same as LibreTranslate’s HTTP API, and the README says translator settings can only be changed in the GUI. We’d recommend TranslateLocally if a person is pasting text. We’d recommend LibreTranslate if the caller needs an HTTP API. They come out level because one is the window and the other is the API. The pages we checked never publish a language count, so “how many languages” stays unanswered.

Argos Translate is the library the API above is made of. The README says it is an offline Python library, a command line, or a GUI, and then says the GUI code is in a separate repository. Models are zip files ending in .argosmodel. PyPI requires Python 3.5 or newer, and describes the engine as OpenNMT’s CTranslate2.

The language list on the README is long, from Arabic through Urdu, and it ends with “and more”. There is no total. A pair you did not install can pivot through one you did, and the same paragraph says that costs some translation quality.

We’d recommend Argos Translate if you are writing the program and want the translation call in process, on the machine, with no HTTP service. We’d recommend LibreTranslate if you want that service already wrapped. The $1,000 figure is for a custom model trained on your data, per language. It is not a price for the library.

Apertium does not try to be Argos or LibreTranslate, which use a neural engine, or TranslateLocally, the Marian and Bergamot desktop app. The README says it uses deterministic linguistic rules rather than statistical or neural models. The FAQ is stricter: neural and statistical tools are not included, and contributions of that kind are not accepted for now. That is a choice, and it is also the reason to skip Apertium when you wanted a neural model.

The wiki says 51 pairs have released versions and are considered stable. Simpleton UI is the documented easy path on Windows and Mac, and it installs pairs from the website, not the in-development ones. Packages exist for Debian, Ubuntu, Fedora, CentOS, OpenSUSE, Windows, and macOS. Installing the core on Windows 10 or newer uses WSL. The install wiki says deep work is easier on Ubuntu or a similar Debian system. An Android app does offline translation and does not yet support every released pair.

We’d recommend Apertium when your language pair is one of those 51 and you want rules you can read. We’d recommend LibreTranslate when you want Argos’s neural engine and the text still has to stay off a vendor’s server. We’d recommend TranslateLocally when you will paste the text into the Marian and Bergamot app rather than use written rules.