Honesty over hype
Every benchmark number ships with a confidence interval. No claim we can't back up with a test.
About · Founded 2023 · v1 launched 2026-05-27
An independent AI research lab. Built in Bratislava. Released to the world.
Mission
We believe the next decade of AI will be defined as much by what models can do with tables of data as by what they can do with text. Most real business decisions still happen on rows and columns — clinical records, financial transactions, sensor logs, lab results. The AI built for this kind of data is years behind the chatbots everyone is using. We're closing the gap, openly.
Zero One Research is an independent AI lab based in Bratislava, Slovakia. We publish our models under an open license — anyone can use them commercially, no strings attached. Our goal is to make AI for tables and spreadsheets as natural to use as AI for text.
How we work
We optimize for three things, in this order: calibration, smallness, and openness. Bigger is not better when smaller works. A point prediction without uncertainty is a guess. A closed-weight model that can't be audited is an externality our customers can't accept. Everything we ship follows from those three commitments.
Small enough to audit
Smaller is better when smaller works. Models you can run on a laptop, fine-tune yourself, and read end-to-end in an afternoon.
Transparency
Upfront about team size, compute budget, and where we still fall short.
Open by default
Open weights. Open research posts. Anyone can build on what we publish — commercially or for fun, no strings attached.
Tables, not chatbots
We chase problems where AI is years behind — tables, records, structured data — not where it's already crowded. New methods, not new branding.
Founder
Independent. Bratislava. Bootstrapped.
Matej founded Zero One Research in 2023 to build the open-weight foundation models that regulated industries can legally deploy on their own data. Before Zero One Research he ran an independent AI consultancy in Slovakia, delivering production machine learning systems to clients across financial services, healthcare, and energy.
He built the full PredictLM v1 stack solo — pretraining (PyTorch + CUDA), distillation, evaluation harness, marketing site, PyPI packages, MCP server, infrastructure. PredictLM v1 shipped on 2026-05-27 as two open-weight models on Hugging Face (Mini 13M, Base 26M, both Apache-2.0), a Python package, and an MCP server for LLM agent integration.
He maintains one of the few public independent implementations of the BarDistribution regression head outside the original TabPFN authors' codebase.
Hiring plan after pre-seed close: research engineer (pretraining scale-out), inference engineer (hosted-tier serving), ML researcher (evaluation pipeline).
Get in touch
We work with companies on pilot deployments, with researchers on collaborations, and with prospective hires on roles. Contact us or follow our research feed.