Lara Mini: How Translated trained an AI model in 45 languages

More than 20 years of work in translation and AI led Translated to a question. What happens when you train an AI model in 45 languages from day one?

The answer is Lara Mini. Translated trained this general-purpose AI model from scratch in 45 languages, including all 24 official languages of the European Union. It ranked first overall among 36 open base models with up to 2 billion active parameters in Translated’s evaluation.

What did Translated learn from training in 45 languages?

Translated built Lara Mini around text from many languages and cultures from the start. Across the full training run, 41% of the tokens came from English data and 36% came from data in the other 44 languages. The rest came from code, math, and instruction-style examples.

To test whether more English would improve results, the team replaced some of the multilingual data with additional English in an experiment. English scores barely moved, while multilingual performance fell significantly. The final model went on to rank second on English and first on multilingual understanding in Translated’s evaluation.

How did Lara Mini rank first in its class?

Translated ran all 36 models through the same 26 benchmarks. English, multilingual understanding, math, and coding each counted for a quarter of the overall score. For the multilingual tests, Translated averaged results across the 44 non-English languages used to train Lara Mini.

Lara Mini scored 57.0 overall, 3.1 points ahead of the runner-up. It achieved that score with 600 million active parameters per token. The model has 7.2 billion parameters in total, but it uses only a subset to process each token. The ranking compares models by how many parameters they actively use.

For the team at Translated, this is a milestone more than two decades in the making. It gives Lara 4 a foundation built around linguistic and cultural diversity from the first day of training.

What will Lara Mini make possible for Lara Translate users?

Translated is using Lara Mini as the base model for offline text translation in the next Lara Translate app. Once the feature is available, you’ll be able to translate text on a plane or anywhere you lose your internet connection. The app will process the translation directly on your phone, so your text will stay on the device.

FAQ

Can I use Lara Mini in Lara Translate today?

Lara Mini is currently unavailable in Lara Translate. The Lara Translate Research page has details on the research release, model access, and the technical report.

What is the difference between Lara Mini and Lara 4?

Lara Mini is the general-purpose base model Translated has released for research. Lara 4 is the broader family of models that will power the next generation of Lara Translate. Its planned offline translation feature will build on Lara Mini.

Will offline translation support all 45 languages?

The 45 languages refer to Lara Mini’s training data. Translated’s research announcement does not specify which languages the offline feature will support at launch.

This article is about

Translated’s Lara Mini research release, what the team learned from training an AI model in 45 languages, and how that work will help bring offline text translation to Lara Translate.