The AI Lab at Poland’s National Information Processing Institute (OPI) has developed PolDense, a new family of language models designed specifically for information retrieval. It is the latest solution created by the OPI research team as part of its efforts to develop advanced artificial intelligence technologies tailored to the Polish language and the needs of domestic users and institutions.
PolDense models have been designed for systems that process large volumes of unstructured data, particularly search engines, chatbots, AI assistants and applications using the increasingly popular Retrieval-Augmented Generation, or RAG, architecture.
The quality of the PolDense models is reflected in their performance in the Polish Information Retrieval Benchmark (PIRB), one of the most important benchmarks for evaluating the effectiveness of Polish-language information retrieval models. PolDense 1B currently ranks first in the public PIRB ranking, achieving the highest average score among all models evaluated.
AI Lab Develops Key Components of Modern AI Systems
A growing number of solutions based on large language models use the RAG approach, which allows them to generate answers not only from the knowledge stored within the model, but primarily from up-to-date documents retrieved in real time and relevant to a particular domain or industry.
The effectiveness of such systems depends largely on the quality of their information retrieval mechanisms. The best results in this field are currently achieved by so-called dense retrievers. These models use deep neural networks to transform queries and documents into compact vector representations, allowing content to be retrieved based on its meaning rather than solely on keyword matching.
This is precisely the type of information retrieval technology currently being developed by the OPI AI Lab. The team is building the PolDense model family for the Polish language, while also working on EuroDense, a model that will support the most widely used European languages.
“The release of the PolDense models is another step towards building Poland’s expertise in artificial intelligence. We are creating open technologies that can be used by researchers, public authorities and businesses to develop modern, efficient and secure AI tools. It is worth emphasising that one of OPI’s new models ranks first in the Polish Information Retrieval Benchmark, outperforming multilingual models such as Llama and BGE,” said Jarosław Protasiewicz, Director of the National Information Processing Institute.
OPI Model Takes the Top Position in the PIRB Ranking
The performance of the PolDense models in the Polish Information Retrieval Benchmark confirms their high quality. PIRB is one of the most important benchmarks used to evaluate the effectiveness of information retrieval models for the Polish language.
PolDense 1B currently occupies first place in the PIRB ranking, recording the highest average score among all evaluated models. It outperforms significantly larger multilingual solutions, including models with nearly 10 billion parameters.
The result confirms that the solutions developed by the OPI AI Lab are among the world’s leading information retrieval technologies for the Polish language. At the same time, they offer a more attractive balance between performance and implementation costs.
“PolDense 1B’s first-place position in the PIRB ranking demonstrates that specialised models designed for the Polish language can set quality standards not only among domestic models, but can also compete effectively with international and multilingual models many times their size. This achievement is the result of many years of research conducted by the OPI AI Lab into modern neural representation models and specialised information retrieval models for Polish,” said Sławomir Dadas, Deputy Head of the AI Lab at OPI.
Six OPI Models Designed for Different Applications
The PolDense family consists of six models based on the ModernBERT architecture and ettin-encoders technology. They range from 17 million to 1 billion parameters, allowing users to select a model suitable both for deployments requiring the highest possible quality and for environments with limited computing resources.
The models can analyse texts of up to 8,192 tokens, enabling them to retain the context of longer documents more effectively and identify the most relevant information with greater accuracy.
The largest model in the family, PolDense 1B, scored 64.11 points in the PIRB benchmark, establishing a new performance level for Polish-language information retrieval. It also outperformed significantly larger multilingual models, including Llama-Embed-Nemotron-8B, which scored 63.73 points, and BGE-Multilingual-Gemma2-9B, which achieved 63.26 points.
The models with 400 million and 150 million parameters offer highly competitive results compared with solutions containing several billion parameters. The smallest variants, with 68 million, 32 million and 17 million parameters, have been designed for deployment on CPUs, mobile devices and edge computing environments.
Open OPI Solutions for Research, Public Administration and Business
The National Information Processing Institute has released the PolDense models as open-source solutions. They can therefore be used to develop custom search mechanisms, RAG systems and tools supporting work with document databases.
PolDense models can not only improve the quality of information retrieval but also reduce implementation costs. Organisations can use smaller and more computationally efficient models without sacrificing a high level of performance.
“PolDense shows that it is possible to develop specialised Polish-language models that not only compete with the world’s largest solutions but also outperform them in many specific tasks. Our objective was to create a family of models suitable for a wide range of applications, from large corporate systems to lightweight solutions operating locally. We are making all the models available free of charge on Hugging Face to support the development of Poland’s AI ecosystem,” said Marek Kozłowski, Head of the AI Lab at the National Information Processing Institute.
“We are delighted with the success of the PolDense models, but we are not resting on our laurels. In the coming months, the OPI AI Lab plans to release EuroDense, which will support information retrieval in nine European languages,” Kozłowski added.
PolDense Models Developed Under the LLMs4EU Project
The PolDense models were developed as part of the Large Language Models for the European Union project, known as LLMs4EU, which is being implemented by the Alliance for Language Technologies European Digital Infrastructure Consortium, or ALT-EDIC.
The project aims to develop and provide artificial intelligence models, tools and services for five key sectors: research, public services, tourism, telecommunications and energy.
It is also intended to encourage public institutions and small and medium-sized enterprises to adopt European AI technologies. LLMs4EU is co-financed by the European Union under the Digital Europe Programme and by Poland’s Ministry of Digital Affairs.
The PolDense models are available free of charge on the Hugging Face platform.
