Nums AI introduces causilo, a tabular foundation model leading tabarena benchmarks

Nums AI has launched Causilo, a pretrained tabular foundation model designed for classification and regression tasks. The model, available via a scikit-learn interface and on Hugging Face, achieves the highest Elo score among single models on TabArena for both tasks. Deployable today for research and evaluation on CUDA or CPU, Causilo requires a separate license for commercial use. Causilo employs in-context learning, storing training data as context rather than updating weights. It processes inputs such as NumPy arrays and pandas DataFrames, handling categorical features and missing values. Version 1.0.1 introduces median and quantile outputs. The model’s architecture includes three phases—refinement, compression, and in-context learning—each detailed in the released code. Causilo outperforms models like TabFM and TabPFN-3 on TabArena, with an Elo score of 1794. It also ranks first in CRPS, R², and RMSE on ScoringBench. Nums AI reports that Caus,ilo is the fastest model for both fit and predict, though TabPFN-3 uses less GPU memory. The model requires Python 3.10–3.12 and PyTorch 2.13 or newer.