Polish AI Models: A Practical Guide

Most mainstream language models treat Polish as a side language. It gets decent coverage, but inflection, long compound sentences, legal phrasing and local cultural references often trip them up. That gap has pushed Polish researchers, universities and volunteer communities to build models trained with Polish text at the centre. Some are large general-purpose assistants, others are compact encoders built for classification and search, and each suits a different kind of job.

Models at a glance

Model	Developer	Type	Size range	Licence	Best for
Bielik	SpeakLeash and ACK Cyfronet AGH	Decoder LLM	1.5B to 11B	Open (Apache 2.0 on most releases)	Polish text generation, RAG, local deployment
PLLuM	HIVE AI consortium led by Wrocław University of Science and Technology	Decoder LLM family	8B to 70B	Open, some non-commercial variants	Public administration, official and legal Polish
Qra	OPI PIB and Gdańsk University of Technology	Decoder LLM	1B, 7B, 13B	Open	Research and Polish fine-tuning bases
Trurl 2	Voicelab	Chat LLM	7B, 13B	Llama 2 licence	Conversational and multilingual tasks
HerBERT	Allegro	Encoder (BERT)	Base and large	Open	Classification, NER, semantic search
papuGaPT2	Community project	Decoder (GPT-2)	Small	Open	Lightweight experiments and teaching

Bielik

Bielik is probably the best-known name in the group. It grew out of SpeakLeash, a community that spent years collecting and cleaning Polish text, working with the Cyfronet supercomputing centre in Kraków, which provided training time on its Helios machine. The family runs from about 1.5 billion up to 11 billion parameters, and the 11B v3 release, published in late 2025, extends coverage to other European languages while keeping Polish as its strongest point.

Features

• Built on depth up-scaling, where a smaller base model is expanded in layers and then trained further
• Instruct versions tuned for chat and task following
• Many quantized builds, so it runs on modest GPUs and even some laptops
• Open weights on Hugging Face

Strengths

• Punches above its size on Polish benchmarks, including cultural knowledge tests where it has beaten several models many times larger
• Strong at Polish writing, rewriting and summarising
• Cheap to host, which matters for small teams and on-premise setups
• A good base for domain tuning; a legal-focused fine-tune has reportedly beaten much larger general models on Polish legal tasks

PLLuM

PLLuM, short for Polish Large Language Model, is the state-backed effort. It began as a consortium of Polish research institutions and continued as the HIVE AI project, commissioned by the Ministry of Digital Affairs. The family spans roughly 8 to 70 billion parameters, so there is a version for quick tasks and a heavier one where accuracy matters more than speed. In 2025 the team released eleven new models under open licences documented to meet EU AI Act requirements.

Features

• Training leans on manually prepared, organic Polish data rather than text generated by other models
• A responsible AI layer with data governance and an output correction and safety filter
• Variants built on different foundations, including Llama and Mixtral-style architectures
• Chat and instruction versions

Strengths

• Comfortable with official, administrative and legal registers
• Clear documentation on data sources and evaluation, which helps with compliance
• Scales from small to very large, so it covers more deployment scenarios than most Polish models
• Backed by a wide group of institutions, which gives it long-term support

Qra

Qra came from a collaboration between the National Information Processing Institute and the Gdańsk University of Technology. Rather than training from scratch, the team continued pretraining existing Llama 2 models on a large Polish corpus, producing 1B, 7B and 13B versions.

Features

• Continued pretraining on Polish text over an existing strong base
• Three sizes for different hardware budgets
• Published as foundation models, not finished assistants

Strengths

• Good starting point for researchers who want to build their own instruction-tuned model
• Noticeably lower perplexity on Polish text than the base models it grew from
• Small sizes are easy to experiment with

Trurl 2

Trurl 2 is a Llama 2 derivative fine-tuned by Voicelab on Polish and English conversational data, released in 7B and 13B sizes. It was one of the earlier Polish-focused chat models to appear, and it still turns up in production prototypes.

Features

• Bilingual Polish and English behaviour
• Instruction and dialogue tuning
• Built on Llama 2, so tooling support is mature

Strengths

• Handy for assistants that switch between Polish and English in the same conversation
• Straightforward to run with common open-source inference stacks
• Useful as a baseline when comparing newer models

HerBERT

HerBERT is a different kind of model. Developed by Allegro, it is a BERT-style encoder, so it does not write text. It reads text and turns it into representations that other systems can use.

Features

• Base and large versions
• Trained on Polish corpora with attention to Polish morphology
• Easy to fine-tune for specific labelling tasks

Strengths

• Excellent for sentiment analysis, named entity recognition and classification
• Fast and light compared with generative models
• Reliable backbone for search and retrieval pipelines where an embedding model is needed

papuGaPT2

papuGaPT2 is a GPT-2 model trained on Polish text by a community team. It is small and dated by current standards, but it remains a friendly way to learn how generative models behave without needing serious hardware.

Features

• Compact decoder architecture
• Open weights

Strengths

• Runs almost anywhere
• Useful for teaching, demos and quick fine-tuning experiments

How they compare with global models

Honest benchmarking matters here. A March 2026 test of twelve models by Oxido, reported in the Polish press, placed Bielik and PLLuM well below the largest global systems on a set of Polish language tasks, with Bielik scoring 6.38 and PLLuM 5.95 against 8.13 for Gemini. Tasks such as reciting famous literary passages exposed the gap.

The builders push back on that framing. These models are not databases of memorised books, they cost a tiny fraction of what frontier systems cost to train, and they are aimed more at business and institutional use than at casual chat. Independent studies point the same way: in a retrieval-augmented generation test on technical documentation, Bielik 11B outperformed a comparable PLLuM model in answer accuracy.

The practical takeaway is to match the tool to the job. For private deployment, data control and Polish-first output, the home-grown models make a lot of sense. For broad reasoning or coding, a global model may still be the stronger pick, and many teams combine both.