OPUS-MT review: free offline models from Helsinki

Using OPUS-MT for Mexico, Brazil and India

The OPUS-MT model for English to Romance languages takes a target code at the start of each sentence, and that code can name a country: es_MX for Mexico, es_CO for Colombia, pt_BR for Brazil, fr_CA for Canada. One model of about 300 MB covers Spanish, Portuguese, French, Italian and Romanian from English.

For India, separate English to Hindi and English to Urdu models are published, each around 300 MB. English to Spanish alone was downloaded about 166,000 times in the past month.

OPUS-MT licenses: Apache 2.0 or CC BY 4.0

OPUS-MT models ship under either Apache 2.0 or CC BY 4.0, and the license is set per model. The 86 newer tc-big models use CC BY 4.0, which requires attribution, and so do some older ones such as English to German. English to Hindi and English to Romance languages use Apache 2.0. Check the model card before putting a model into a commercial site.

Running OPUS-MT on your own server

A classic model weighs about 300 MB and a tc-big model about 465 MB. You load them in Python with the Hugging Face Transformers library, and translation then runs offline. Each model covers one direction, so a site publishing in four languages loads four models.

Without code, the same group publishes OPUS-CAT, an MIT-licensed app for Windows, macOS and Linux that runs OPUS-MT models locally and plugs into Trados Studio and memoQ.

No text leaves your machine, and no usage terms touch your content. There is no service, warranty or support either: this is a university research project. The models checked here were last updated between August 2023 and February 2024.

Direct language pairs in OPUS-MT

Many models translate between two languages without going through English, such as French to Spanish or French to German. Argos Translate, by contrast, routes every pair through English except Portuguese and Spanish. For an English-language site, this difference does not matter.

OPUS-MT strengths and limitations