Nums AI has released Causilo, a pretrained tabular foundation model for classification and regression that achieves the highest Elo score among single models on TabArena. The model utilizes an in-context learning approach where training rows are stored as context rather than updating weights, and it is available with Apache-2.0 code and pretrained weights on Hugging Face.

  • Causilo tops TabArena with an overall Elo of 1794, outperforming single models like Google's TabFM and LG AI Research's EXAONE Tabular.
  • On ScoringBench, it ranks first by CRPS, R², and RMSE across 101 datasets.
  • The architecture uses cross-attention to keep computational cost linear in feature count, with a default ensemble of 8 members sharing weights.
  • Version 1.0.1 adds median and quantile outputs for regression tasks.
  • In speed tests on an H100 GPU, Causilo was the fastest for both fitting and prediction, though TabPFN-3 used significantly less memory.

The model is deployable for research and evaluation on CUDA or CPU, while commercial use requires a separate license from Nums AI.