Poland’s AI takes its next leap

Bielik is evolving from a Poland-focused language model into a multilingual tool with European ambitions. Its creators are preparing a new model with more than 100 billion parameters.

Sebastian Kondracki, twórca Bielik
Sebastian Kondracki, head of the Bielik project, has set his sights high. He wants Polish AI to become a key European technology. He has a concrete plan. Photo: Bielik press materials
Loading the Elevenlabs Text to Speech AudioNative Player...

The new version of Bielik is designed not only to hold conversations but also to carry out tasks. Its creators have also unveiled models for automating decisions. Their next goal is to build an AI model with more than 100 billion parameters. “Europe should not rely on a single player,” Sebastian Kondracki, head of the SpeakLeash Foundation, tells XYZ.

Bielik is moving beyond its origins as a project focused primarily on Polish. The SpeakLeash Foundation and the Academic Computer Centre Cyfronet AGH have just announced the launch of Bielik 11B v3.1. The model, which supported 32 languages in its previous version, now handles 56, including Chinese, Japanese, Korean and Arabic. At the same time, it has gained capabilities that allow it to form part of a system that carries out a sequence of actions, rather than simply answering questions.

The launch actually encompasses several products. Alongside the main model are four Bielik-One models for classifying information and supporting decisions, as well as Bielik DFlash, which is designed to reduce the time needed to generate answers. All are to be released with their weights under open science principles. For businesses, this means an opportunity to build solutions on their own infrastructure instead of relying on an external service for every operation.

From language assistant to task performer

Version 3.1 underwent additional training on 500 billion tokens, the units of text used to train a model. The expanded training data was intended to take Bielik beyond its previous range of European languages without sacrificing its Polish-language capabilities or grasp of the local context. The new Bielik supports all official EU languages, as well as Armenian, Azerbaijani and Georgian, among others.

“We are not building Bielik just so that Poland has its own language model. We are building it because we believe Polish technology can set standards, rather than merely follow them. The launch of Bielik 3.1 is another important step toward that ambition. We want Bielik to become the model of first choice in Europe and, in time, beyond it,” Sebastian Kondracki, president of the SpeakLeash Foundation, tells XYZ.

Tests point to progress in language comprehension. In Belebele, a benchmark that assesses reading comprehension through questions about texts, the score rose by 8 percentage points compared with version 3.0. In INCLUDE, a test of knowledge and reasoning in 44 languages that takes regional contexts into account, the improvement was 5 percentage points.

The second change, however, may be much more significant for commercial applications. Bielik 3.1 combines function calling, structured output and reasoning. Behind these technical terms are three different abilities: using functions and tools made available to the model, returning information in a format that other systems can understand, and reasoning through more complex instructions.

Good to know

The history of Bielik

The story of Bielik began in 2022, with the creation of SpeakLeash, an initiative also known as Spichlerz. It was founded by Sebastian Kondracki, and its first goal was to create a vast, open collection of Polish texts for training artificial intelligence. Researchers, programmers, technology company employees and students became involved, initially contributing largely on a voluntary, unpaid basis.

A breakthrough came through cooperation with the Academic Computer Centre Cyfronet AGH, which provided access to its Helios and Athena supercomputers. This made it possible to train language models on a scale beyond the reach of most Polish research teams.

The first Bielik 7B debuted in April 2024, with 7 billion parameters and an architecture based on that of France’s Mistral. Bielik 11B v2, with 11 billion parameters, followed in August that year. The model was made available on open terms, allowing others to develop it further and run it on their own infrastructure.

In 2025, the Bielik family began to grow. Its creators introduced, among other things, version 2.5 of the 11B model and smaller v3-generation models with 1.5 billion and 4.5 billion parameters. The project increasingly focused on efficiency, adapting models to specific uses and reducing their computing requirements.

In 2026, Bielik entered another stage. Version 3.0 expanded its multilingual and reasoning capabilities.

A test where the right answer is not enough

The creators tested these capabilities using τ²-bench, a benchmark that simulates multistep customer service in industries including airlines, retail and telecommunications. The tasks do not stop at finding the correct answer. The model must achieve a specified outcome — for example, use a tool to check a reservation or change a service — while following established rules.

Depending on the scenario, Bielik 3.1 scored 16–45 percentage points higher than version 3.0. This was the largest improvement in agentic capabilities reported in the announcement.

“Bielik 3.1 also uses reasoning mode when calling tools. This is especially important for more complex tasks, where the model has to plan and execute a specific sequence of actions and choose the right tools along the way. Combined with function calling and structured output, this allows Bielik to handle multistep tasks more effectively, as the τ²-bench results also show,” says Krzysztof Ociepa, head of AI model training at the SpeakLeash Foundation.

Four new models built to evaluate, not write

The Bielik-One family has debuted alongside the main model. It comprises four decision models of different sizes, allowing users to choose according to the available computing power and the task at hand. Unlike traditional generative models, they were not primarily designed to create natural-language responses. Their role is to assess the data provided and assign it to one of a set of categories.

Bielik-One is designed to return probabilities for each possible answer. For example, if a system analyzes messages in a company inbox, it can assess how likely a message is to concern a complaint, a question about an offer or an administrative issue. The company can then use that assessment to set rules for automatically routing cases. If a result does not meet the required confidence threshold, the message is sent for further review.

“In many business processes, we do not need an extensive response from a model, but a specific result — for example, assigning a request to the right category or assessing it on a particular scale. Bielik-One has been adapted precisely for tasks like these. Instead of generating text, it indicates the probability of each possible answer, allowing a system to use the result according to the rules set for a given process,” says Krzysztof Wróbel, lead researcher at the SpeakLeash Foundation.

The creators also point to uses in sentiment analysis, preliminary assessment of applications, checking compliance with procedures and evaluating content generated by other AI systems.

The third part of the launch is Bielik 11B v3.1 DFlash. It is a so-called draft model that works with a larger main model. Instead of waiting for Bielik to generate every successive part of an answer on its own, the lighter model proposes how the text should continue. The main model then checks these proposals and accepts those it considers correct. The mechanism is known as speculative decoding.

One million downloads in a month

New model versions are not the only measure of progress. The head of SpeakLeash points to interest in Bielik among users of open repositories, as well as cooperation with institutions and major companies.

“Bielik is now the second most downloaded European model family on Hugging Face. In our record month of August, we exceeded one million downloads. I am even happier about adoption we can document. PFR runs the ‘Start with Polish AI’ campaign, in which companies talk about their implementations. It also launched a free certification program, which around 10,000 people have already completed. Thanks to our cooperation with InPost, Bielik has been integrated into an app used by more than 15 million people,” Kondracki says.

He names cooperation with the Polish Armed Forces and NVIDIA among the project’s most important achievements so far.

“The result is Bielik Minitron, a model that is one-third smaller and up to 50% faster than the 11B version. I am also particularly pleased that our base model has become the foundation for other models on the market. That is the best proof that open weights work,” he adds.

Expert's perspective

Do local language models make sense?

Bielik 11B v3.1 now understands 56 languages, and its Polish tokenizer requires more than 40% fewer tokens for Polish text than the popular foreign model Llama. That means less computing and a lower bill. Most importantly, however, a local model can be run in an organization’s own server room, keeping the data of a public office, hospital or bank inside the building. With the AI Act and inspections by the Commission for the Development and Security of Artificial Intelligence, that is not an extra — it is a ticket to regulated sectors.

Bielik’s challenge is not technology, but money. The model is free, its creators report 4.7 million downloads, and a foundation develops it with the help of partners, patrons and Cyfronet’s supercomputers. That is enough to build an excellent model, but keeping pace with the rest of the world is difficult. Mistral, which Bielik’s creators want to compete with, raised EUR 3 billion in the third quarter, while the Canadian government is buying a stake in its domestic AI company Cohere. At the same time, according to Menlo Ventures, 89% of global corporate spending on open models now goes to Chinese models. If the Polish model cannot find a way to make money, companies will choose a cheaper and faster competitor.

I do not think Bielik needs to make money from the model itself. The model is a horse: it has strength but does not know where to go. The money is in the harness — the implementation: connecting it to an organization’s systems and knowledge, managing permissions and costs, demonstrating compliance and taking responsibility for results. I see three paths. First, the state can be a customer, not just a patron. The French tax administration is currently turning to Mistral models. Second, there can be a paid enterprise layer: versions with guarantees, support and certification, while the model remains free for everyone. Third, an ecosystem of Polish implementation companies can build products on Bielik and share revenue with its creators.

At allclouds.pl, we are building such a harness: SAIE, in which the model can be swapped out and data stays with the customer. The stronger the Polish model, the better for everyone implementing AI in Poland. That is why I am rooting for Bielik and hope that, alongside details of the model with more than 100 billion parameters, we will also hear at the Bielik Summit on November 26 how its creators intend to make money from it.

American APIs, Chinese open models and a European third way

A particular vision for the market underpins Bielik’s expansion. SpeakLeash does not want to compete solely as a provider of a ready-made assistant. It intends to distribute models on which other teams can build their own products. Kondracki sees this approach as an alternative to the model prevalent in the US, where access to closed services is sold commercially.

“The United States is primarily about commercial, closed APIs. China has chosen a different route and is conquering the world with open weights. On OpenRouter, Chinese models already account for most of the tokens processed. On Hugging Face, meanwhile, new models are increasingly being built on Chinese base models. It is a very effective distribution method,” Kondracki says.

The European approach is also intended to account for obligations under the AI Act. The project’s head says the team will develop solutions to support training-data documentation, transparency and content labeling.

“This is precisely where we want to break in: open weights with full AI Act support, from showing what we train on to content-labeling technologies. The second pillar is the expertise needed for adoption. Together with European technology companies and startups, we want to teach people how to fine-tune models, build new ones on our base models and calibrate them. The aim is to give European companies and institutions a model they understand and control,” Kondracki explains.

More than 100 billion parameters: the next Bielik will be built from scratch

The biggest change is still to come. The Bielik Summit, scheduled for November 26, is expected to bring details of work on a model with more than 100 billion parameters. For comparison, the latest publicly unveiled Bielik is called 11B, referring to its approximately 11 billion parameters. The new project is therefore set to be more than nine times larger by that measure.

According to Kondracki, it is intended to be a multimodal model. It is also being built from scratch on an original architecture, with support from Polish private-sector companies and institutions. According to the announcement, the team has been developing the project for six months.

“We have always said that our path is not so much about the model itself as about developing the ability to build the next one. Each version of ‘little Bielik,’ which is more like a compact car than a huge truck packed with electronics, was a planned stage in acquiring those skills. We started with text, then added reasoning capabilities […] and agentic abilities, which are very strong in version 3.1 announced today. On top of that, we have worked continuously with business, including InPost,” Kondracki says.

An original architecture and a scale of more than 100 billion parameters are intended as a step toward a European model. Kondracki points to France’s Mistral as a benchmark. In his view, Europe needs another provider of open technology in this class.

“We have the expertise to start building a European model. One that not only speaks Polish, English or Chinese well, but can also code and has agentic capabilities useful in everything from business to e-commerce. Today, virtually one company in Europe is building open models of this scale: France’s Mistral, and it is the only one that features prominently in rankings such as the Intelligence Index. No open science project has yet crossed the 100 billion-parameter threshold. Europe should not rely on a single player, which is why we want to join it with our own model, built from scratch on an original architecture,” Kondracki says.

Key Takeaways

  1. A multilingual model with agentic capabilities. Poland’s Bielik is becoming a multilingual model and gaining features needed to build AI agents. The SpeakLeash Foundation and Cyfronet AGH have unveiled Bielik 11B v3.1, which supports 56 languages, up from 32. The new version can also use external tools, plan sequences of actions and pass results to other IT systems. In tests simulating multistep customer service, it scored 16–45 percentage points higher than its predecessor. This is an important step toward applying Polish technology.
  2. A growing family of specialized models. Alongside the main model, its creators have unveiled four Bielik-One variants for classifying information and supporting decisions, as well as Bielik DFlash, designed to speed up response generation. Potential applications include customer service, document analysis, assessment of requests and automatic routing to the appropriate departments. All the models are to be released with their weights under open science principles, allowing companies to develop their own AI systems without having to depend on external service providers to operate them.
  3. European ambitions. A much larger model is in development to compete in the European AI market. SpeakLeash Foundation president Sebastian Kondracki has announced plans for a multimodal model with more than 100 billion parameters, based on an original architecture. Details are to be presented at the Bielik Summit on November 26. The strategy involves developing open technology, supporting European companies in adopting AI and ensuring compliance with EU regulations. Bielik has already worked with NVIDIA, InPost and the Polish Armed Forces. According to Kondracki, its models exceeded one million downloads on Hugging Face in their record month. The goal is to create a European alternative to technologies developed in the United States and China.