AI Blog: Bielik — The Polish Open-Source Language Model
May 6, 2025
Bielik is a Polish open-source large language model developed by the SpeakLeash Foundation with the support of computing resources provided by the Academic Computer Center Cyfronet AGH (one of the fastest supercomputing centers in Poland as of 2025).
The project aims to deliver a model specifically tailored to the Polish language and cultural context. Bielik is trained entirely on Polish data, including legal texts, the Polish constitution, Wikipedia, and Project Gutenberg materials available under open licenses.
The first version of Bielik was released in April 2024, followed by version 2 on August 28, 2024. As of May 6, 2025, the team released version 3 in two sizes (1.5B and 4.5B parameters), along with version 2.5 with 11B parameters. All are available under the Apache 2.0 license.
Earlier versions were based on the Mistral-7B architecture. From version 3 onward, Bielik uses the Qwen 2.5 base model and a custom tokenizer specifically optimized for the Polish language.
You can run Bielik online or even locally on your own GPU. This makes it suitable for educational, business, and personal use cases — especially for developers and researchers working with the Polish language.
The team behind Bielik consists of 50–80 contributors, including data curators, model trainers, and open-source tool developers. One of the main coordinators of the project is Sebastian Kondracki.
In addition to the official website, Bielik is also available on the Hugging Face platform, where users can freely test and download the model.
The developers are also working on a companion project called Sójka, which allows content filtering based on user-defined categories such as violence or profanity — particularly useful in educational settings.
Bielik was awarded Technology of the Year by Money.pl in May 2025. It has been well-received by the Polish AI community for its openness, local relevance, and the ability to run entirely offline without sending data to external servers.
In 2024, the project received an AI Spotlight Award at the GOSIM conference and a Special Recognition at the Masters&Robots event.
The model Bielik-7B-Instruct-v0.1 ranked highly on the Polish OpenLLM Leaderboard, especially in tasks like information retrieval and text classification.
Bielik proves that open, sovereign AI development is not only possible, but thriving — even in smaller language communities.