Translated launches Lara Mini: Globally Ranked No 1 in its Class

Trained on 45 languages and cultures, including all 24 official EU languages, the new European AI model ranks #1 in its class worldwide, challenging the belief that training on more languages comes at the cost of intelligence.

Translated has announced Lara Mini, a new general-purpose large language model built on a premise that linguistic and cultural diversity is not an obstacle to artificial intelligence, but a potential source of greater intelligence.

While most AI models are trained predominantly on English, Lara Mini was trained from scratch on 6 trillion tokens of data in 45 languages and cultures, including all 24 official languages of the European Union.

Translated’s experiments tested the conventional trade-off directly and found that replacing multilingual training data with additional English yielded almost no improvement in English performance, while significantly reducing multilingual capabilities.

With just 0.6 billion active parameters, Lara Mini achieved the highest score of all open base models evaluated with up to 2 billion active parameters on a combined assessment of English, multilingual understanding, mathematics, and coding, including models from Google, Meta, and Alibaba. It outperforms models trained on up to six times more data. This result was made possible by combining exceptional quality, built on over 20 years of translation data collected by Translated, with an original approach to bringing these elements together.

Marco Trombetti, Co-founder and CEO of Translated, comments:

“For too long, AI has treated the diversity of human languages as a problem to solve. We believe the opposite: different languages and cultures offer different perspectives on reality. That diversity can make intelligence richer, not weaker. Lara Mini is our first step toward proving it at scale.” 

Lara Mini is the first step in Translated’s ambition to develop world-leading, efficient AI models that learn from humanity’s full linguistic and cultural diversity rather than primarily from one language. Translated will now investigate whether this approach could produce models competitive with frontier AI systems at a fraction of today’s training-compute cost. The ambition is to demonstrate this possibility on a large scale.

The model weights, technical report, and a new multilingual reasoning benchmark covering 110 language variants are being released for research and non-commercial use.

Lara Mini is part of Lara 4, a new family of foundation models that will power the next generation of Lara Translate, Translated’s AI translation service. By learning from many languages and cultures from the outset, rather than treating multilingual ability as an add-on, Translated expects these models to capture meaning, context and cultural nuance more accurately. Combined with Lara Translate’s specialised training, this should bring faster, higher-quality translation.

Translated has acknowledged and thanked CINECA for its support and for providing access to the Leonardo supercomputer used in the research leading to Lara Mini.

Research and results: https://laratranslate.com/research

 

Related Post: Translated will bring real-time Holy Mass from St Peter’s Basilica in 60 languages

1602 UCAD Translated
author avatar
Trish Stevens Head of Content
Trish is the Head of Content for In the Channel Media Group. [email protected]
Share by Email
Facebook
Twitter
Whatsapp
LinkedIn

Related Articles

Featured

Read our latest magazine