Latvian researchers have published the world’s first artificial intelligence model that equally represents Latvian and other European languages. In terms of language support and safety, the new model successfully competes with technology giants from the US and China. The open-source model TildeOpen enables the creation of diverse AI services that work excellently in Latvian and other underrepresented languages. This unique model, commissioned by the European Commission and developed by the Latvian company Tilde, is freely available to anyone interested.
The development of TildeOpen became possible after Tilde won the European Commission’s Large AI Grand Challenge, granting them 12 months of access to Europe’s most powerful supercomputer - LUMI in Finland. Two million supercomputer hours were allocated for training the model. The model was created by Tilde’s research lab team, which includes nine PhDs and extensive international research experience.
TildeOpen has 30 billion parameters and was trained on hundreds of billions of words from all European languages. For Latvian alone, 29 billion words were used- the largest known dataset ever applied to an AI model in the language.
"When using popular AI models in Latvian, we often see grammatical mistakes and awkward phrasing, because they are primarily trained on English. TildeOpen was designed differently - all European languages are treated equally, resulting in much higher quality output," explains Tilde CEO Artūrs Vasiļevskis.
Early testing shows that TildeOpen performs 41% better in Latvian than the popular LLaMA-3 model and 24% better than GPT-4o. When adapted for translation tasks, TildeOpen translates texts significantly better than GPT-4.1 models of comparable size and performs on par with the supermodel behind ChatGPT, which is 60 times larger.
The model is developed in full compliance with the EU AI Act and supports all 24 official EU languages as well as Ukrainian and other candidate country languages. Unlike global models that typically operate in US or Asian data centres, TildeOpen can be deployed on local servers or in European cloud infrastructure, ensuring data storage within EU borders in line with European data protection standards.
TildeOpen is freely available on the Hugging Face platform and can be used by public institutions, businesses, researchers, and students to build custom AI solutions - from information processing and customer service to education and cultural projects.
"Our goal was to prove that it is possible to develop world-class AI technology right here in Latvia. TildeOpen is a major step toward digital equality for Latvian and other smaller languages," emphasises Vasiļevskis.
