Nums AI has launched Causilo, a tabular foundation model that utilizes in-context learning. The model operates by storing training rows as context and performing predictions in a single forward pass without updating pretrained weights. It was trained exclusively on synthetic data.
The architecture consists of three phases: refinement, compression, and in-context learning. It processes features in groups of three using cross-attention to maintain linear costs relative to feature counts. The model compresses rows into 512-dimensional vectors and uses a 12-layer prediction block.
Causilo achieved an Elo rating of 1794 in the TabArena benchmark, surpassing models such as TabFM, EXAONE Tabular, and TabPFN-3. It also ranked first in ScoringBench across CRPS, R-squared, and RMSE metrics based on evaluations across 101 datasets.
The model is currently available for research and evaluation on CPU or CUDA, requiring Python 3.10 to 3.12 and PyTorch 2.13 or newer. Commercial use and hosted API access require a separate license from Nums AI.
