Small Language Models
Training Polish language models from scratch. We study scaling laws, tokenizers, data mixtures and distillation to understand how much capability a compact model can learn.
Explore programme →
FABRYKA AI RESEARCH
INDEPENDENT ARTIFICIAL INTELLIGENCE RESEARCH LABORATORY · WARSAW, POLAND
We build models.
Then put them
to work.
We study how language models learn, how to make them more efficient, and how to evaluate them reliably — with a particular focus on Polish-language AI.
MODELS · SYSTEMS · EVIDENCE
FROM QUESTION TO PUBLIC ARTIFACT
Research → Experiments → Evidence → Applications
01 / RESEARCH PROGRAMMES
Three lines.
One research agenda.
How much quality can we achieve with a given budget of data, training and inference — and how can we measure it reliably? These three programmes connect our current work and planned studies.
Efficient Inference
Making language models practical on commodity GPUs. We study quantization, speculative decoding, including DFlash, routing and the trade-offs between throughput, latency and quality.
Explore programme →Polish AI Evaluation & Data
Building the data and measurements that Polish AI needs. This programme connects Polish DynaWord, CodeSOTA benchmarks, OCR evaluation and methodologies for comparing Polish LLMs.
Explore programme →02 / SELECTED RESEARCH OUTPUTS
Work you can inspect.
Explore public model weights, data and benchmark evidence. Each linked artifact carries its own documentation and scope.
Pollock Mini LM · 125M
A compact language-model release. Model card, weights and usage information are available in the public repository.
Polish DynaWord
A public Polish-language dataset supporting our work on language-model training and data composition.
CodeSOTA
Model comparisons and benchmark evidence. Individual entries identify their sources and evaluation context.
03 / OPEN RESEARCH
Follow the work.
Inspect the evidence.
Our research connects questions, baselines, experiments and artifacts. We use Fabryka Track to record training runs, compare configurations and investigate results.
Dated experiment write-ups are preserved in our research notes. Track provides the tools for recording experiments; access to individual runs depends on the project.
04 / RESEARCH INFRASTRUCTURE
Tools for doing
the research.
Training, measurement and serving are part of the same experimental workflow. We build the infrastructure that helps us investigate each step.
CodeSOTA
Public infrastructure for model comparisons, benchmark results and evaluation evidence.
Explore CodeSOTA ↗Fabryka Track
Experiment tracking for training runs, metrics, logs, checkpoints and configuration comparisons.
Explore Track ↗Compute laboratory
GPU systems for training and inference experiments, with hardware and workload conditions documented alongside results.
Inference research →05 / PEOPLE
A lab is the people
doing the work.
Meet the Fabryka AI team: researchers, engineers and contributors working across models, data, evaluation, agents and infrastructure.
Meet the Fabryka AI team ↓- 15
- team members & contributors
- 2
- advisers
Kacper Wikieł
Language models, training infrastructure, model optimization and evaluation.
CO-FOUNDER / RESEARCH & ENGINEERING
Arkadiusz Słota
Datasets, evaluations, training infrastructure and systems engineering.
Piotr Zientara
Founder of Xfaang. Builds autonomous AI agents and LLM products, connecting data, evaluation and infrastructure.
Piotr Styła
Agent teams, web and mobile applications, recruitment and onboarding.
Kuba
Theoretical physics, computer vision for oncology, and generative and agent systems.
Kamil Kaczmarek
Bots, microservices and production tools, from idea to deployment.
Konrad Talik
Long-running agent workflows and ML strategies informed by data exploration and visualization.
Piotr Rybarczyk
Data pipelines, systems engineering and technical planning.
Maciej Sawicki
Evaluation tooling, reproducible experiments and Polish byte-level BPE tokenizers.
Radomir Mastalerz
Research tooling, experiment documentation and presentations.
Dawid Majewski
Training infrastructure, developer tools and the Fabryka Track interface.
Adam Skrodzki
Small-model training experiments, model-based evaluation and benchmark methodology.
Łukasz Czerwiński
Software engineering and contributions to the Fabryka Track interface.
Paweł Puzio
Polish tokenizers, technical reviews and research planning.
@bartoszkobylinski
Contributions to research discussions, planning and review.
RESEARCH & PRACTICE
Our advisers
Supporting the team with research methodology, evaluation and domain expertise.
06 / TECHNOLOGY TRANSFER
Research,
put to work.
Selected technologies developed or evaluated at Fabryka become production systems. Our inference platform makes open models available through an API, with tools for chat, routing and token-level analysis.
The platform also serves models developed by other teams. Availability through our API is separate from authorship of a model.
Explore Fabryka API →07 / ABOUT FABRYKA AI
Independent research.
Built in Poland.
Fabryka AI is an independent artificial intelligence research laboratory in Warsaw. We build datasets, train models from scratch and develop methods for efficient inference and reliable evaluation.