Bielik AI: Polish LLM language model
In a world dominated by global giants such as OpenAI, Google, and Anthropic, Poland has taken its own bold step toward digital independence. Bielik AI—the first open large language model created in Poland—is becoming a symbol of national innovation. Based on open-source principles, it processes the Polish language with native-level precision while ensuring full control over data and infrastructure. In November 2025, it gained new computing power and global recognition, being added to NVIDIA’s model catalog.
Table of contents
.
Why Do We Need Bielik?
At the outset, we may ask ourselves why Bielki is needed at all, given that ChatGPT, Grok, Claude, and Gemini already exist, and without a large budget, it will be difficult to create something better. Additionally, there are also open-source projects (Llama, DeepSeek, GPT-OSS) that anyone can download onto their own computer and use free of charge, without limitations, paying only for electricity.
However, our own model has quite a few additional advantages. In the case of models developed by others, we do not know exactly what data they were trained on. In the case of Bielik, we know exactly what data was used to train the model. Bielik was trained on texts written in Polish, and we have access to all the texts used to train it.
Community, tools, and investment in the future.
The main advantage of Bielik is the entire community of enthusiasts that has formed around this project. The Discord server is a true treasure trove of knowledge and is very useful for anyone learning to create language models, RAG systems, or machine learning.
.
If in the future someone in Poland decides that they need their own language model and is able to allocate several hundred million for it, carrying out such a project will be easier thanks to Bielik already existing. Bielik also serves as a proof of concept, which could potentially encourage other people to invest in AI.
Open-source AI models are not always entirely open.
In the case of traditional open-source programs, anyone who knows programming can modify such a program, because the source code in which the given program was written is publicly available. Closed-source code consists of zeros and ones, so figuring out how a given program works is very difficult.
Closed-source code may be paid or made available free of charge. Sometimes the user has no access to the code at all; this is how online applications and programs work, where the user sends a request to the server and all operations run on the software provider’s server.
In the case of AI models, a significant part of the program consists of embedding matrices and arrays containing parameters and weights. This means that even if a given model is free and available for download, modifying it is not easy.
.
Project Genesis
Bielik was born in 2024 on the initiative of the SpeakLeash foundation (also known as Spichlerz). Its founders, led by Sebastian Kondracki, decided to create a Polish language model based on open principles. The Academic Computer Centre CYFRONET AGH in Kraków provided support by making the Helios and Athena supercomputers available.
The first version of the model was released on March 24, 2024, and just a month later it became available to a wider group of users. Bielik was trained on vast Polish datasets: from legal documents and statutes, through Project Gutenberg, to the Polish Wikipedia. As a result, it learned to understand nuances of language, style, and culture that global models struggle with.
By 2025, the team consisted of 10–15 people, five of whom were strictly focused on training models. Subsequent versions—2.0 (August 2024) and 3.0 (May 2025)—brought significant improvements. Version 2.5 premiered in Paris under the patronage of the Ministry of Digital Affairs.
.
How many people use Bielik?
In October 2025, the Bielik-11B-v2.3-Instruct model was downloaded 120 thousand times.In total, since its inception, it has been downloaded one million times.
Bielik Technical Specifications.
The first version of Bielik was based on the French Mistral-7B model and had 11 billion parameters. Although it was smaller than its American counterparts (GPT-3 had 175 billion), it performed excellently on tasks in Polish.
Version 2.3 (Instruct) retained the scale of 11 billion parameters, while introducing a “decoder-only” architecture based on transformers. The latest iterations (v3) bring two smaller variants—1.5 and 4.5 billion parameters—based on the Chinese Qwen 2.5 and optimized for the Polish language. All models are available on the Hugging Face platform and the bielik.ai website under the Apache 2.0 license.
With the Jan.ai application, users can run the model locally without the need to use external clouds. This guarantees data privacy and security, which is of great importance to the public, medical, and financial sectors.
Capabilities and Applications of Bielik, from the Office to Healthcare
Bielik does not compete with giants in scale, but in adaptability. Trained on 90% Polish texts, it performs exceptionally well in local contexts and tasks, such as:
- Text processing – generating summaries, sentiment analysis, classification of documents and messages;
- Data security – local deployment protects information from leaks, which is crucial in administration and medicine;
- Business integration – automation of reports, analysis of regulations, support in customer service and documentation.
The Ministry of Digital Affairs highlights Bielik as an example of an effective tool for Polish businesses that want to use AI without becoming dependent on foreign providers.
Additionally, the Sójka project—an extension of Bielik—introduces content filters, such as profanity blocking and ethical context analysis, thereby increasing the system’s social responsibility.
The Polish Model on the World Stage
In a short time, Bielik gained recognition both in Poland and abroad. In 2024, the SpeakLeash Foundation received the “AI Spotlight” distinction at the GOSIM conference, as well as the special Masters&Robots award. In May 2025, the model won in the Technology of the Year category at the Money.pl awards ceremony.
On the Polish OpenLLM Leaderboard, Bielik-7B-Instruct-v0.1 shines in classification and information retrieval tasks. The project community already numbers nearly 4,000 people, making it one of the largest open AI projects in Central Europe.
Bielik in the NVIDIA Model Catalog
June 2025 brought a breakthrough event – Bielik AI was added to the NVIDIA NIM APIs and NVIDIA NGC catalogs, becoming part of the global ecosystem of models supported by the American giant. The Bielik-11B-v2.3-Instruct version is now available to developers on the build.nvidia.com platform, with full support for GPUs from the Hopper and Blackwell series.
Integration with NVIDIA infrastructure makes it possible to run Bielik in an environment compatible with the OpenAI API, facilitating deployment in enterprise applications. The model has been optimized for GPUs, which increases inference performance and reduces energy consumption.
According to NVIDIA’s documentation, Bielik supports LoRA techniques, offering precise text processing in cultural, legal, and technical contexts. This proves that Polish technology can be both local and global—combining sovereignty with global standards.
.
Bielik AI and Billennium Collaboration
The collaboration between the SpeakLeash Foundation and Billennium represents one of the key steps in the development of Bielik AI. Billennium, a Polish technology company with over 20 years of experience and offices in Poland, Germany, India, and Malaysia, specializes in creating IT solutions for business and public administration. Thanks to this partnership, Bielik gains practical applications in commercial and public projects, such as document automation, citizen services, and the analysis of official data. The joint efforts of both entities strengthen the idea of sovereign Polish artificial intelligence, developed locally, with respect for privacy and full control over data.
Bielik Development Plans
In autumn 2025, Bielik gained access to the F.I.N. supercomputer at the AI Factory of Beyond.pl in Poznań. Built on the NVIDIA DGX SuperPOD platform with B200 GPUs, it will enable the training of even larger models and the development of enterprise versions.
A new initiative called Gniazdo aims to create an AI ecosystem for Polish companies, accelerating the development of specialized domain models. The plans also include:
- Bielik Citizen, collecting multimodal data (text, images, audio);
- AI watchdog, i.e. a system for monitoring ethical deployments.
Cooperation with Cyfronet and AGH remains the backbone of the project. At the same time, the Polish AI ecosystem is being strengthened by EU initiatives such as PIAST (Poznań, over PLN 200 million) and Gaia (Kraków, PLN 300 million), which support the development of domestic language models.
The Significance of Bielik AI for Poland.
In the era of the global struggle for data, Bielik represents a real step toward digital sovereignty. It is not only a technological tool, but also a symbol of independence. Thanks to it, Poland gains the ability to build its own AI solutionswithout having to use commercial foreign models.
As Sebastian Kondracki emphasizes, the real challenge lies not only in computing power, but also in responsibility and ethics. Bielik brings these two worlds together—science, the open-source community, and democratic values.
.
Summary
Bielik AI is more than just a technology project. It is proof that collaboration among scientists, enthusiasts, and public institutions can produce tangible results. In just two years, Poland created a model that gained international recognition and became part of NVIDIA’s global ecosystem.
Bielik is not the “Polish ChatGPT” – it is something more important: a tool for building technological independence and an inspiration for future AI projects in Europe.
Try it on bielik.ai or build.nvidia.com and see for yourself that the future of Polish artificial intelligence has wings of its own.