Skip to content
Open Source

Apache-2.0 by default. Commercial use, no asterisks.

v1 model weights, training code, evaluation harnesses, and benchmarks are published under Apache-2.0. Future premium models may be hosted-only — we say so on each release. Commercial deployment, fine-tuning, and redistribution of open releases are all fine.

GitHub

github.com/matej-01RAI

Training scripts, evaluation harnesses, MCP server, migration tools.

  • predictlm — main Python package (training + inference)
  • predictlm-mcp — MCP server for LLM agent integration
  • tabpfn-to-predictlm — migration CLI for TabPFN users
View on GitHub

Hugging Face

huggingface.co/zerooneresearch

Model weights, model cards, and reproducible evaluations.

  • predictlm-mini-13m — the compact deployment checkpoint
  • predictlm-base-26m — the reference architecture
View on Hugging Face

Cite this work

If you use PredictLM in academic or commercial work, please cite the model and the architectural-experiments writeup:

@misc{predictlm2026,
  author       = {Zero One Research},
  title        = {PredictLM: open-weight tabular foundation models},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/zerooneresearch/predictlm-mini-13m}
}