Polish PolDense Model Tops Information Retrieval Ranking

Poland's Information Processing Institute (OPI) has released the PolDense family of language models, designed specifically for information retrieval in Polish-language texts. Its flagship variant, PolDense 1B, had the least trouble of all, taking first place in the Polish Information Retrieval Benchmark (PIRB) ranking, ahead of models many times its size.

PolDense isn't another large generative model, it's a family of so-called dense retrievers, models that convert text into numerical vectors used to find the most relevant document passages. This kind of technology underpins RAG (Retrieval Augmented Generation) mechanisms, which let chatbots and AI assistants draw on current, specific data instead of relying solely on knowledge baked into a model's parameters during training.

PIRB Ranking Results

The Polish Information Retrieval Benchmark, created earlier by OPI's AI Lab itself, covers several dozen Polish-language information retrieval tasks, from legal documents to medical and business texts. In this benchmark, PolDense 1B scored 64.11 points, putting it in first place ahead of Nvidia's Llama-Embed-Nemotron-8B (63.73 points) and BGE-Multilingual-Gemma2-9B (63.26 points) from the Chinese lab BAAI. Both of those models have nine times more parameters than their Polish rival.

The smaller model's edge over much larger competitors comes down to language specialization. Multilingual models have to split their capacity across dozens of languages at once, while PolDense was trained solely around the specifics of Polish, its inflection and specialized vocabulary.

Six Variants for Different Uses

The PolDense family includes six models ranging in size from 17 million to 1 billion parameters. The smallest variants are meant for lightweight, local deployments, for example on mobile devices or systems with limited computing resources, while the largest, PolDense 1B, targets large enterprise systems. All variants handle texts up to 8,192 tokens, allowing long documents to be processed without splitting them into small chunks.

The team emphasizes that the models are intended primarily for RAG systems, chatbots and AI assistants, as well as internal document search engines at companies and public institutions. This is an area where retrieval quality translates directly into the accuracy of answers generated by larger language models.

Funding and Open License

The project was developed under the LLMs4EU initiative, funded by the European Union and Poland's Ministry of Digital Affairs. All PolDense models have been made freely available on the Hugging Face platform under an open license, meaning they can be used by commercial companies as well as research institutions and public administration without licensing fees.

Releasing the PolDense models is another step in building Polish competencies in artificial intelligence. We are creating open technologies that can be used by researchers, public administration and entrepreneurs to build modern, effective and secure AI tools - Dr. Jarosław Protasiewicz, Director of the Information Processing Institute

Dr. Marek Kozłowski, head of AI Lab at OPI, points to the project's goal in the context of its range of applications. The team wanted to deliver models scalable from large corporate deployments to simple, local installations, while keeping the technology openly accessible to the entire Polish AI ecosystem.

Our goal was to develop a family of models that meets the needs of very different applications, from large corporate systems to lightweight deployments running locally. We are making all the models available for free on the Hugging Face platform to support the development of the Polish AI ecosystem - Dr. Marek Kozłowski, Head of AI Lab, Information Processing Institute

What's Next

OPI's AI Lab team says it will continue work on specialized information retrieval models. The next step is expected to be EuroDense, a model supporting retrieval in nine European languages, which OPI says will launch in the coming months. It extends the approach proven with PolDense to a European scale, under the same LLMs4EU initiative.

PolDense's success fits into a broader trend of building specialized, smaller language models for Polish, alongside existing projects like Bielik and PLLuM. Unlike those, which focus on text generation, PolDense addresses a different, though equally important, layer of AI infrastructure, information search and indexing, which underpins practical RAG deployments at companies and institutions.