Slovenia Launches GaMS AI Model

By Suad Seferi · Aug 10, 2026

Slovenia has developed an open-source large language model designed specifically for the Slovene language, giving researchers, companies and public institutions a locally deployable alternative to relying entirely on global AI platforms. The University of Ljubljana's Faculty of Computer and Information Science presented GaMS, a family of generative AI models adapted for Slovene and intended for use in areas including healthcare, industry, cultural heritage and software development. The models can be downloaded and installed locally, while a public chatbot is also available through the project's website. For Slovenia, the project addresses a problem familiar to many smaller European countries. The largest AI systems are developed using enormous datasets dominated by English and other widely spoken languages. Languages with smaller digital footprints often receive less representation in those datasets. GaMS shows one way that smaller language communities can respond without attempting to build a frontier AI model entirely from scratch. Building AI around the Slovene language GaMS was developed through the PoVeJMo project at the University of Ljubljana, coordinated by researchers working on language resources and technologies for Slovene. One of the project's central challenges was collecting enough high-quality material. Slovene has around two million speakers, making the available pool of digital text much smaller than that available for English. Researchers therefore invited Slovene speakers to contribute their own writing through the povejmo.si platform. Training material was also contributed by organisations including Slovenia's National and University Library, the newspaper Dnevnik and the Slovenian Press Agency. Rather than creating the underlying technology from zero, researchers adapted existing open models. A research paper published in March describes GaMS3-12B, a 12-billion-parameter model developed through continued pre-training of Google's Gemma 3 model followed by supervised fine-tuning. The training mixture contained around 140 billion Slovene, English, Bosnian, Serbian and Croatian tokens, followed by more than 200,000 Slovene and English instruction examples. The inclusion of neighbouring South Slavic languages is particularly relevant for Southeast Europe. It shows how linguistic resources from related languages can contribute to the development of AI for a language with a smaller digital footprint. GaMS improves on its base model in Slovene According to the research team, GaMS3-12B performed better than the original 12-billion-parameter Gemma 3 model across the Slovene evaluation scenarios used in the study. Researchers tested the model on Slovene benchmarks, English-to-Slovene translation and a Slovene LLM arena where users compare responses produced by different models. The paper reports that GaMS3-12B achieved a win rate above 60 percent in the Slovene arena and performed comparably to the much larger commercial GPT-4o model in that particular evaluation. That comparison should be interpreted carefully. It does not mean GaMS performs at the same level as frontier commercial systems across every task or language. The more useful result is that focused training significantly improved an existing open model's performance for Slovene. That is an important distinction for countries that do not have the computing infrastructure or financial resources required to train the world's largest AI systems. Slovenian companies are already testing GaMS The project has also moved beyond academic evaluation. Four Slovenian companies have adapted GaMS for different applications. Healthcare technology company Better has worked with the model in medical and clinical environments. Špica has explored industrial speech recognition, including situations involving background noise, dialects and multilingual workers. Semantika has worked on applications for museums and cultural heritage, while XLAB has explored the model's use in generating descriptions and documentation related to computer infrastructure. Local deployment is another important part of the project. The GaMS models are openly available for organisations, researchers and developers to install in their own environments and adapt for specialised applications. The project's open-access catalogue currently includes several versions, including GaMS3-12B, GaMS3-12B-Instruct and other instruction-tuned variants. Running a language model locally can also allow organisations to keep sensitive information within their own infrastructure rather than automatically sending data to external AI services. The University of Ljubljana has highlighted this data-control aspect as one of the reasons for developing domestic AI infrastructure. Why Slovenia's GaMS matters for the Western Balkans Slovenia's experience raises a larger question for the region: who will build AI that properly understands Europe's smaller languages? Macedonian, Albanian, Bosnian, Serbian, Croatian and Montenegrin operate in the same global AI market, where the quantity of available training material and investment is heavily concentrated around larger languages. Commercial AI platforms increasingly work across many of these languages, but language support alone does not guarantee the same quality of understanding, specialised vocabulary or locally relevant knowledge available in English. GaMS offers a different approach. Instead of trying to compete directly with global technology companies by training enormous foundation models, universities and public institutions can adapt capable open models using carefully prepared local datasets. For parts of the Western Balkans, regional cooperation could make that approach more practical. The GaMS researchers have already used Bosnian, Croatian and Serbian material alongside Slovene during training. Shared language resources, evaluation frameworks and computing infrastructure could similarly help countries with limited datasets build…