Kacper Wikieł
CEO / Research
Responsible for research, models and sales. Creator of polish-dynaword and the “LLM od zera” course, with experience in AI/ML and enterprise deployments.
Fabryka AI / Our mission
Polish AI.
Open by design.
We are building an ecosystem for hosting open models on sovereign Polish infrastructure.
We are a Polish AI lab developing open-weight models. Our goal is best-in-class models, open data and benchmarks, and infrastructure located in Poland.
01 / Principles
Mission before profit.
Fabryka AI is a for-profit company, but profit is not our overriding goal.
When profit and mission conflict, we will choose the mission. We will not close what we have released openly. We use our profits to fund research and compute. We will not sell the company to anyone who would shut down that openness.
02 / Why we do this
Our own models. Our own expertise.
We want Polish companies to have access to AI developed and operated in Poland. We are building a local alternative to sending data and per-request payments to providers abroad.
Many applications depend on doing one task well: understanding a document, extracting data, or having an agent call the right tool. That is why we are starting with small, specialised models. Our goal is to match or outperform leading general-purpose models on specific tasks, at a lower cost.
Open weights alone are not enough. Data, training processes and evaluations are also needed so others can check the results and reproduce how they were achieved.
03 / Development plan
Scaling one step at a time.
0.3–1 billion parameters
Models for specific tasks: document understanding, data extraction and tool calling. Each follows the same path: a public benchmark, data preparation, fine-tuning, open weights and a report, then an API.
2–4 billion parameters
Broader tasks built on a proven process. Each stage is intended to validate our methods and generate revenue to support operations and further research.
A Polish general-purpose model below 30 billion parameters
A larger training run using the data and infrastructure built along the way. Our ambition is a model among the best in its class, with publicly verifiable results.
04 / Open development
Models, data, benchmarks — all in the open.
We aim to publish models with reports and benchmarks, and data with documentation of its provenance and licences. Openness should enable independent verification, reproducible results and collaboration with researchers.
We are pursuing this through polish-dynaword, ZusWaveBench and the Open PL LLM Leaderboard. Data and evaluations are the foundation for future models, not an afterthought at launch.
CodeSOTA helps us understand which tasks people need models for. Public comparisons should guide what we train — and then show whether we have met our goal.
05 / From research to product
An API today. Our own models next.
The Bielik API is already live. API access to PLLuM models is planned.
We will build dedicated endpoints for models adapted to the needs of companies and institutions. We plan to support fine-tuning on customers’ own data and host the resulting models on Polish infrastructure.
Our business model starts with a token-based API running on our own GPUs. The next stage is custom models, including for customers in regulated sectors: fine-tuned on their data and deployed on our infrastructure or theirs. Developing our own base model is intended to pave the way for further model adaptation services.
06 / Team
Research and engineering under one roof.
Arkadiusz Słota
Co-founder / Engineering
Responsible for engineering, infrastructure and product. Builds datasets, evaluations and training infrastructure, drawing on experience with industrial and agent-based systems.
07 / Our goal
Poland as a contributor to AI.
We want Poland to become a significant contributor to AI built through open development. We are building models, infrastructure and expertise that other Polish products can use. The value created by AI research and deployment should also be created here.