{"id":11265,"date":"2025-02-25T15:00:00","date_gmt":"2025-02-25T14:00:00","guid":{"rendered":"https:\/\/haimagazine.com\/uncategorized\/new-polish-language-model-pllum\/"},"modified":"2025-06-26T15:10:25","modified_gmt":"2025-06-26T13:10:25","slug":"new-polish-language-model-pllum","status":"publish","type":"post","link":"https:\/\/haimagazine.com\/en\/ai-news\/new-polish-language-model-pllum\/","title":{"rendered":"\ud83d\udd12 New Polish language model: PLLuM"},"content":{"rendered":"

On February 24, 2025, a conference was held at the Ministry of Digital Affairs where the family of Polish language models, PLLuM (Polish Large Language Model<\/em>), was officially presented. The conference marked the culmination of the first phase of a project developed by a consortium of six scientific institutions, led by the Wroc\u0142aw University of Technology under the auspices of the government (with a budget of 14.5 million zlotys). The goal was to build comprehensive Polish LLMs in order to support public administration and academic communities in the digitization process, as well as Polish business in terms of innovation.<\/p>

\"\"Stages of Development of PLLuM, Ministry of Digital Affairs<\/p>

The PLLuM family utilizes from 8 to 70 billion parameters depending on the specific model. The smaller ones are designed for simple tasks, such as creating a bot to manage websites. Larger models are better suited for areas where contextual coherence in understanding the Polish language is needed, for example in scientific research or processing official documentation. All versions are based on ethically sourced data\u2014<\/em>under licenses, the amended copyright law, and EU regulations. <\/strong>Additionally, scientific models (which aren’t approved for commercial use) use publicly available data sets such as Common Crawl. This makes them even more efficient.<\/p>

\u201cTogether with the Bielik model, they can promote artificial intelligence developed in Poland, supporting each other in a better training process and further data acquisition and opening, which is necessary for #AIMadeInPoland to keep improving\u2014for public administration, business and society.\u201d <\/em>Ministry of Digital Affairs.<\/em><\/p><\/blockquote>

An interesting moment of the conference was the confirmation that one of the PLLuM models (8×7-nc-chat) outperforms in competency tests<\/mark><\/a> benchmarks of counterparts like GPT-4-turbo or DeepSeek R1-Llama-70B. This confirms that the success of the latter was not accidental and that massive investments aren’t necessary to create efficient LLMs that greatly understand the nuances of our language and culture. Similarly, in the area of security, the Polish model (12b-chat) performs well when it comes to resistance to disruptions (e.g., imprecise user queries).<\/p>

The Ministry has announced that for 2025, 19 million PLN has been secured for further development of the project and research. The HIVE consortium, which is in charge of implementing the project (this time led by NASK), has expanded by two new partners: the Central Informatics Center and the Academic Computer Center Cyfronet AGH. Currently, the inclusion of private companies in the consortium is legally limited, but efforts are underway to change the regulations to allow this.<\/p>

The following implementations of models in the public administration are planned for 2025: <\/strong><\/p>