Nums AI Releases Causilo, a Tabular Foundation Model With Apache 2.0 Code

In mid-September 2026, Nums AI shipped Causilo as a pretrained tabular foundation model for classification and regression. PyPI shows causilo 1.0.1 around September 13–14, with secondary coverage landing September 15–16. The Python package and GitHub repository ship under Apache 2.0; Hugging Face weights at nums-ai/causilo use Causilo License v1.0, which permits non-commercial research and evaluation while requiring separate licenses for commercial, production, and hosted API use.

Project materials and secondary write-ups describe the fit path as in-context learning: calling fit stores training rows as context, and prediction runs in one forward pass so query rows attend to labeled training rows without modifying that context. Inputs accept NumPy arrays and pandas DataFrames with categoricals and missing values. Classification supports up to 10 classes. Version 1.0.1 exposes median and quantile regression outputs from 999 native quantiles.

Benchmarks and caveats

Nums AI reports TabArena Full single-model Elo of 1792.9 overall (classification 1771.8, regression 2032.6) using the official 51-dataset pipeline, with a maintainer re-run cited at overall Elo 1794. Secondary coverage notes that when system ensembles are included, the same re-run placed Causilo third of 88 overall, and that top Elo confidence intervals overlap. Treat leaderboard claims as project-reported until independent production audits appear. Requirements listed in the README are Python 3.10–3.12 and PyTorch 2.13+.

Primary sources are the nums-ai/causilo GitHub README and Apache LICENSE, the Hugging Face model card and weight license, PyPI causilo 1.0.1, and mid-September secondary reports that mirror those primaries.