Fabryka AI started with a thought that was probably arrogant: I thought I could build a model better than Bielik.
The Bielik team had done something worth respecting. They had brought people together, built models, and given others something concrete to learn from. Seeing their work made me want to try my own ideas.
It also made me impatient. I wanted to see that technical talent matched with bigger ambitions: raise capital, build a research company, and try to compete with teams like Mistral.
I couldn’t know whether that would work. But I wanted to find out how far we could go.
Having opinions about someone else’s model is easy. Eventually, you have to get compute and put your own ideas to the test.
So I did.
Give it a goal. Give it compute.
When I got access to an H100, I became interested in another question: how much of the work could an AI agent do?
I wanted to push the /goal feature to its limits. Could I ask an agent to train a model, give it access to compute, and have it get the job done? Where would it get stuck? What would I still need to understand and decide myself?
That question brought together two things I wanted to explore: building models and discovering how much further AI tools could help a small team reach.
The experiment was a way of testing the distance between an idea and its execution. What could someone attempt if they had capable tools, access to hardware, and the willingness to learn?
I wanted to explore that distance firsthand.
People started asking different questions.
Around me, I began noticing something that became just as interesting as the models.
People who had been using APIs and building applications started wanting to understand the technology underneath them.
“Which model should I use?” became “Why does this one work better?”
Then came: “What happens if we change the data?” “Could a different tokenizer help?” “Can we train something better ourselves?”
Developers were beginning to do research. They were reading papers, designing experiments, comparing results, and developing ideas of their own.
That progression stayed with me. Once someone becomes curious enough to investigate how a model works, they need somewhere to take that curiosity: compute, tools, collaborators, and time.
I wanted to build a place where they could keep going.
For the love of the craft.
There is something deeply satisfying about understanding a difficult system.
You have a hypothesis. You change something. You run an experiment. The result forces you to think again. Sometimes an improvement survives closer examination. Sometimes it disappears.
Our own work includes both. In one experiment, an apparent improvement on a small evaluation vanished when we tested it on a larger sample. That result belongs in the record too. Knowing that an approach failed can save someone else from spending another week on it.
I see people around Fabryka becoming absorbed in these questions: how to make a model learn more from its data, how to fit more capability into a smaller system, how to make inference faster, how to measure whether an improvement is real.
That is the culture I want to build around. People who care about the technology enough to understand it deeply and make it better.
Built in Poland, with room to aim higher.
ElevenLabs changed my sense of what was possible for Polish founders and researchers. Their example made building original AI technology for a global audience feel tangible.
I want us to take that possibility seriously.
Poland is where we are building Fabryka. It is where many of our relationships and collaborations begin. Polish-language research matters to people in our community and to customers, and we are proud to contribute to it.
Our ambition extends to the underlying technology: models, training methods, data, evaluation, inference, and the products those advances make possible.
Small models give us a practical place to investigate difficult questions and compete within the resources we have. They also have useful applications of their own. What we learn at this scale should help us decide what to attempt next.
I want our ambition to grow with our capabilities—and to pursue the customers, collaborators, and investment that can support it.
The work is producing results.
As of September 2026, our GoLLeM-v5 128M ranks fourth by composite score on the Glint Tiny-ML Leaderboard.
We built and released Polish DynaWord, a dataset we are proud of. Other researchers have started citing our work. Something that began with our own questions is becoming useful to people pursuing theirs.
I also discussed building Fabryka on Wycinki z przyszłości, episode 69, published on 26 June 2026: access to compute, funding an open research lab, and how people can move from following the work to contributing.
There have been smaller moments of recognition too. In a conversation with Remigiusz Kinas on 99 Twarzy AI, Karol Stryja connected the idea of a Polish AI research and optimization lab with my ambitions. It was encouraging to hear that direction recognized outside our own conversations.
These are early achievements. They give us concrete reasons to keep building.
The initial belief that we could do better now has experiments, public artifacts, and competitive results behind it.
Make things people want. Keep the research going.
Curiosity needs time. Experiments need compute. People need to earn a living.
I want Fabryka to build products people value enough to pay for, and use that revenue to fund further research.
Our inference platform is part of that effort. Serving useful models creates demanding engineering problems: latency, throughput, reliability, and cost. Improvements in those areas can make the product better while giving us new questions to investigate.
The other projects support the same work. CodeSOTA helps us compare models. Fabryka Track helps us understand experiments. Our datasets and training runs help us investigate how models learn.
We are building towards a company where useful products support research, and research gives us better products to offer. Commercial success would give us more time, more compute, and more freedom to pursue difficult ideas.
That is a business I want to build for the long term.
See how far we can take it.
Fabryka began with my competitive instinct. With Arek and the people contributing alongside us, it has become a shared effort.
We have models on a leaderboard, public data, tools, and research that others are beginning to cite. We also have many questions we cannot answer yet.
What keeps me going is seeing people become capable of work they once thought was beyond them—and knowing how much more we could attempt with the right resources.
I still want to build better models.
Now I want to build the company where we can keep doing it.
