Free and Open Source Machine Translation API, entirely self-hosted. Unlike other APIs, it doesn't rely on proprietary providers such as Google or Azure to perform translations. Instead, its translation engine is powered by the open source [Argos Translate](https://github.com/argosopentech/argos-translate) library.
By default LibreTranslate will install support for all available languages. To only load certain languages and reduce startup time, you can use the **--load-only** argument:
```bash
libretranslate --load-only en,es,fr
```
Check also all other [arguments](#settings--flags) below.
You can use hardware acceleration to speed up translations on a GPU machine with CUDA 12.4.1 and [nvidia-docker](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html) installed.
Arguments passed to the process or set via environment variables are split into two kinds.
- Settings or runtime flags used to toggle specific runmodes or disable parts of the application. These act as toggle when added or removed.
- Configuration parameters to set various limits and configure the application. These require a parameter to be passed to function, if removed the default parameters are used.
| --require-api-key-origin | Require use of an API key for programmatic access to the API, unless the request origin matches this domain | `No restrictions on domain origin` | LT_REQUIRE_API_KEY_ORIGIN |
| --require-api-key-secret | Require use of an API key for programmatic access to the API, unless the client also sends a secret match | `No secrets required` | LT_REQUIRE_API_KEY_SECRET |
| --host | Set host to bind the server to | `127.0.0.1` | LT_HOST |
| --port | Set port to bind the server to | `5000` | LT_PORT |
| --char-limit | Set character limit | `No limit` | LT_CHAR_LIMIT |
| --req-limit | Set maximum number of requests per minute per client (outside of limits set by api keys) | `No limit` | LT_REQ_LIMIT |
| --req-limit-storage | Storage URI to use for request limit data storage. See [Flask Limiter](https://flask-limiter.readthedocs.io/en/stable/configuration.html) | `memory://` | LT_REQ_LIMIT_STORAGE |
| --req-time-cost | Considers a time cost (in seconds) for request limiting purposes. If a request takes 10 seconds and this value is set to 5, the request cost is either 2 or the actual request cost (whichever is greater). | `No time cost` | LT_REQ_TIME_COST |
| --batch-limit | Set maximum number of texts to translate in a batch request | `No limit` | LT_BATCH_LIMIT |
| --ga-id | Enable Google Analytics on the API client page by providing an ID | `Empty (no tracking)` | LT_GA_ID |
| --frontend-language-source | Set frontend default language - source | `auto` | LT_FRONTEND_LANGUAGE_SOURCE |
| --frontend-language-target | Set frontend default language - target | `locale` (match site's locale) | LT_FRONTEND_LANGUAGE_TARGET |
| --api-keys-db-path | Use a specific path inside the container for the local database. Can be absolute or relative | `db/api_keys.db` | LT_API_KEYS_DB_PATH |
| --api-keys-remote | Use this remote endpoint to query for valid API keys instead of using the local database | `Empty (use local db instead)` | LT_API_KEYS_REMOTE |
| --get-api-key-link | Show a link in the UI where to direct users to get an API key | `Empty (no link shown on web ui)` | LT_GET_API_KEY_LINK |
| --shared-storage | Shared storage URI to use for multi-process data sharing (e.g. when using gunicorn) | `memory://` | LT_SHARED_STORAGE |
| --secondary | Mark this instance as a secondary instance to avoid conflicts with the primary node in multi-node setups | `Primary node` | LT_SECONDARY |
| --load-only | Set available languages | `Empty (use all from argostranslate)` | LT_LOAD_ONLY |
| --threads | Set number of threads | `4` | LT_THREADS |
| --metrics-auth-token | Protect the /metrics endpoint by allowing only clients that have a valid Authorization Bearer token | `Empty (no auth required)` | LT_METRICS_AUTH_TOKEN |
- Each argument has an equivalent environment variable that can be used instead. The env. variables overwrite the default values but have lower priority than the command arguments and are particularly useful if used with Docker. The environment variable names are the upper-snake-case of the equivalent command argument's name with a `LT` prefix.
- To configure requirement for api key to use, set `--req-limit` to `0` and add the `--api-keys` flag. Requests made without a proper api key will be rejected.
- Setting `--update-models` will update models regardless of whether updates are available or not.
See [Medium article by JM Robles](https://jmrobles.medium.com/libretranslate-your-own-translation-service-on-kubernetes-b46c3e1af630) and the improved [k8s.yaml](https://github.com/LibreTranslate/LibreTranslate/blob/main/k8s.yaml) by @rasos.
LibreTranslate supports per-user limit quotas, e.g. you can issue API keys to users so that they can enjoy higher requests limits per minute (if you also set `--req-limit`). By default all users are rate-limited based on `--req-limit`, but passing an optional `api_key` parameter to the REST endpoints allows a user to enjoy higher request limits. You can also specify different character limits that bypass the default `--char-limit` value on a per-key basis.
To use API keys simply start LibreTranslate with the `--api-keys` option. If you modified the API keys database path with the option `--api-keys-db-path`, you must specify the path with the same argument flag when using the `ltmanage keys` command.
LibreTranslate has Prometheus [exporter](https://prometheus.io/docs/instrumenting/exporters/) capabilities when you pass the `--metrics` argument at startup (disabled by default). When metrics are enabled, a `/metrics` endpoint is mounted on the instance:
You can use the [official discourse translator plugin](https://github.com/discourse/discourse-translator) to translate [Discourse](https://discourse.org) topics with LibreTranslate. To install it simply modify `/var/discourse/containers/app.yml`:
Then issue `./launcher rebuild app`. From the Discourse's admin panel then select "LibreTranslate" as a translation provider and set the relevant endpoint configurations.
See it in action on this [page](https://community.libretranslate.com/t/have-you-considered-adding-the-libretranslate-discourse-translator-multilingual-to-this-forum/766).
- [LibreTranslator](https://codeberg.org/BeoCode/LibreTranslator) is an Android app [available on the Play Store](https://play.google.com/store/apps/details?id=de.beowulf.libretranslater) and [in the F-Droid store](https://f-droid.org/packages/de.beowulf.libretranslater/) that uses the LibreTranslate API.
- [LiTranslate](https://github.com/viktorkalyniuk/LiTranslate-iOS) is an iOS app [available on the App Store](https://apps.apple.com/us/app/litranslate/id1644385339) that uses the LibreTranslate API.
- [minbrowser](https://minbrowser.org/) is a web browser with [integrated LibreTranslate support](https://github.com/argosopentech/argos-translate/discussions/158#discussioncomment-1141551).
Most of the training data is from [Opus](http://opus.nlpl.eu/), which is an open source parallel corpus. Check also [NLLU](https://nllu.libretranslate.com)
The LibreTranslate Web UI is available in all the languages for which LibreTranslate can translate to. It can also (roughly) [translate itself!](https://github.com/LibreTranslate/LibreTranslate/blob/main/scripts/update_locales.py) Some languages might not appear in the UI since they haven't been reviewed by a human yet. You can enable all languages by turning on `--debug` mode.
In short, yes, [but only if you buy an API key](https://portal.libretranslate.com). You can always run LibreTranslate for free on your own server of course.
### Some translations on libretranslate.com are different than the self-hosted ones. Why?
By default language models are loaded from the [argos-index](https://github.com/argosopentech/argospm-index). Sometimes we deploy models on libretranslate.com that haven't been added to the argos-index yet, such as those converted from OPUS ([thread](https://community.libretranslate.com/t/opus-mt-language-models-port-thread/757))
Yes, here are config examples for Apache2 and Caddy that redirect a subdomain (with HTTPS certificate) to LibreTranslate running on a docker at localhost.
You can remove `127.0.0.1` on the above command if you want to be able to access it from `domain.tld:5000`, in addition to `subdomain.domain.tld` (this can be helpful to determine if there is an issue with Apache2 or the docker container).
To get a HTTPS subdomain certificate, install `certbot` (snap), run `sudo certbot certonly --manual --preferred-challenges dns` and enter your information (with `subdomain.domain.tld` as the domain). Add a DNS TXT record with your domain registrar when asked. This will save your certificate and key to `/etc/letsencrypt/live/{subdomain.domain.tld}/`. Alternatively, comment the SSL lines out if you don't want to use HTTPS.
- Train a new language model using [Locomotive](https://github.com/LibreTranslate/Locomotive). For example, we want to train improved neural networks for [German](https://community.libretranslate.com/t/help-wanted-improve-en-de-translation/935) and many other languages.
- Can you beat the performance of our language models? Train a new one and let's compare it. To submit your model make a post on the [community forum](https://community.libretranslate.com/) with a link to download your .argosmodel file and some sample text that your model has translated.
- Pick an [issue](https://github.com/LibreTranslate/LibreTranslate/issues) to work on.