Nums AI has dropped Causilo, a pretrained tabular foundation model that claims the top spot on TabArena with an Elo score of 1794. This beats out heavyweights like Google Research’s TabFM and LG AI Research’s EXAONE Tabular in single-model evaluations. The model handles both classification and regression tasks using an in-context learning approach. Instead of updating weights during training, Causilo stores training rows as context and predicts query rows in a single forward pass. It accepts NumPy arrays or pandas DataFrames, handling missing values and categorical features natively. For regression, version 1.0.1 now supports median and quantile outputs based on 999 native quantiles.
The architecture is distinct. Nums AI splits the network into refinement, compression, and in-context learning phases. Features are grouped in sets of three, embedded with learned sine and cosine frequencies. A row stage allows feature groups to interact through four latent tokens using cross-attention. This design keeps computational cost linear relative to the feature count, avoiding the quadratic explosion typical of full self-attention. An eight-member ensemble shares weights but cycles through different normalization strategies, boosting robustness without massive compute overhead. Training relied entirely on synthetic data, meaning no TabArena datasets leaked into the pretraining process.
Performance metrics look strong, but context matters. While Causilo leads among single models, a maintainer re-run placed it third overall when system ensembles were included. The confidence intervals at the top overlap significantly, suggesting the lead over TabFM is narrow. On ScoringBench, however, Causilo ranked first by CRPS, R², and RMSE across 101 datasets. Speed tests on an H100 GPU showed Causilo was faster than competitors in both fit and predict phases, though TabPFN-3 used less memory. This efficiency could make it attractive for rapid prototyping where latency is critical.
Deploying Causilo requires caution. The code is Apache-2.0 licensed, but the pretrained weights are for research and evaluation only. Commercial use, production deployment, or hosted APIs require a separate license from Nums AI. It runs on CUDA or CPU with Python 3.10 to 3.12. The shift toward licensing weights while keeping code open mirrors a growing trend in AI, where the real value lies in the trained parameters rather than the architecture itself. Developers should weigh the benchmark gains against the legal friction before integrating it into commercial pipelines.
