OPEN AI. SHARED AMBITION.POLAND → THE WORLD
We build AI
in Poland.
In public.
A place where models, data and research take shape, and people come together to build AI.
We want “making models in Poland” to lead naturally to Fabryka. We combine open research, community and infrastructure into one lab that keeps growing.
The target operating model of Fabryka.
MANIFEST / SEPTEMBER 2026The direction for 2027. Goals we want to reach together.
01 / AMBITION
Poland needs a place
that regularly produces
world-class open AI.
A model is one of the results. The lasting advantage will be the ability to create the next ones: from data and evaluations through training and deployment to new research questions.
We see room here for many teams. We want to grow the Polish ecosystem together with researchers, open-model builders, universities and companies.
02 / ONE LAB, FIVE PROGRAMMES
Fabryka makes things.
Each one strengthens the next.
- 01RESULT → CHECKPOINT
FABRYKA / MODELS
Models for real work.
Polish, reasoning, code and tool use. Regular checkpoints, open weights and documented capabilities.
- 02RESULT → DATASET
FABRYKA / DATA
Data you can trace.
Corpora with documented provenance and licence, community contributions and verified synthetic data. Test sets kept separate from training.
- 03RESULT → BENCHMARK
FABRYKA / EVALS
Progress you can check.
A public methodology, comparable protocols and a protected test portion. Results by category, together with costs and limitations.
- 04RESULT → SHARED WORK
FABRYKA / COMMUNITY
From a first contribution to research.
Concrete tasks, visible authorship and research mentoring. We want to open the way to compute, fellowships and joint publications.
- 05RESULT → THE NEXT EXPERIMENT
FABRYKA / COMPUTE
Infrastructure drives research.
Our target model is to reinvest part of the inference revenue in GPU resources, experiments and open results.
03 / FABRYKA MODEL PROGRAM — PROPOSAL
In Polish.
At a world-class
level.
The goal: the best open model under 15B parameters for work in Polish.
The first direction is FABRYKA-1: an 8–14B-class model, Polish-first, with strong general abilities. The name and scope describe a research plan; the model is not announced here as released.
RESEARCH DIRECTION
<15B
AMBITION / QUALITY PER UNIT OF COMPUTE
- Polish language
- Reasoning
- Code and STEM
- Tool calling
- Documents
- RAG and long context
The reference points will be Polish and global models of the same class. Comparable tests and licence terms will decide the choice of base.
How will we choose the base model?
Planned bake-off: 5–8 candidates on 2–5 thousand Polish tasks. Preliminary weights: Polish language 15%, knowledge and culture 10%, reasoning 15%, coding 15%, agents and tools 15%, documents 10%, long context and RAG 10%, instruction following 10%.
After the base is chosen: small experiments with continued pretraining, SFT and learning from preferences or rewards. We scale only after showing an improvement over the baseline.
What will count as proof of progress?
Ten capability groups: Polish, knowledge, reasoning, code, mathematics and STEM, agents, documents, long context, business tasks and safety. Publishing results requires stating the model revisions, the protocol, the coverage and the limitations.
Protected tests stay out of training. Incomplete measurements earn no place in the ranking. External closed benchmarks serve for final evaluation, not for tuning the model.
04 / THE FIRST 90 DAYS OF THE PROGRAMME
Small experiments.
Visible progress.
A proposed sequence from the programme launch. Later stages depend on results, available compute and the team.
DAYS 01–30
Build a reference point.
The PL Eval v0.1 specification, a comparison of base models and a first version of the dataset. Opening the task list to contributors.
TO PUBLISH
Methodology + baseline + tasksDAYS 31–60
Test the recipe.
Choosing the base, small CPT/SFT experiments and regression analysis. Weekly research notes and the first joint Open Lab session.
TO PUBLISH
Experiments + data + findingsDAYS 61–90
Release the result.
If quality and licences allow: a checkpoint, a model card and an evaluation. Preparing the API and the GGUF, MLX and FP8 formats where they are supported.
TO PUBLISH
Weights + results + how to use them
Every week: a development log. Every 2–4 weeks we want to publish a useful result — a model, dataset, eval, tool or research note.
05 / OPEN FABRYKA
Come in.
Build something with us.
A good eval, a reproduced experiment or a corrected example is a real contribution to the model.
We want every contribution to have an author, a description of its use and a history. The planned path leads from a first task to independent research.
CONTRIBUTORFELLOWRESEARCHER
01 / DATA AND EVALS
Add the knowledge that is missing.
Create examples, check facts, compare answers and build tests for your field.
02 / RESEARCH AND ENGINEERING
Try an idea in practice.
Reproduce results, optimise inference, develop tools and write up experiments.
03 / PARTNERS
Share resources.
Let’s talk about compute, data with clear usage rights, research mentoring and joint projects.
Compute credits, mini-grants and fellowships are elements of the planned programme. The rules and the availability of funding require a separate announcement.
06 / HORIZON: MID-2027
Big ambition.
Explicit criteria.
- 1,000
active contributors
A community goal. We want to count activity by accepted contributions in the last 90 days.
- <15B
a globally competitive model
The ambition: top results in most of the declared categories of work in Polish, at a comparable scale.
- Open
research process
Public artifacts, a regular log and authorship. Successes as well as failed experiments.
These are goals for 2027, not current results or a guarantee of a deadline. We will judge them by published results.
ALREADY TODAY / SEE THE STARTING POINT
Research programmes ↗Publications and results ↗Inference platform ↗
BUILD AI HERE.