Causilo

Causilo is a tabular foundation model for classification and regression, pretrained on about 36 million synthetic tables. It takes the rows you already know the answer for as context and predicts the rest in a single call, with no per-task training run or hyperparameter search. Missing values and categorical columns are accepted as they are. It suits research, benchmarking and testing on tabular data without building a training pipeline.

Validation: 95.1% accuracy on 1,000 held-out rows of the Covertype dataset (UCI) with 50,000 context rows, and R-squared 0.878 on California housing with 10,000 context rows, with no training or tuning. The sample notebook reproduces both figures. Latency on ml.g6e.xlarge for a request of 1,000 query rows with 54 feature columns: about 0.7 s with 1,000 context rows and about 10 s with 50,000.

In the Causilo Technical Report (https://arxiv.org/abs/2609.22866 ), Causilo reaches 1785 Elo on TabArena (51 datasets) at a median 0.10 seconds per 1,000 test rows under the benchmark's timing protocol, which places it on the performance-efficiency Pareto frontier. Its row stages attend through a fixed number of summary tokens, so their attention cost grows linearly with the number of columns.

Known limitations: when training and test rows come from different groups (grouped splits), Causilo ranks behind several conventional supervised models. Not intended for medical diagnosis.

This package runs Causilo inside your own AWS account as a SageMaker real-time endpoint or batch transform job, with network isolation: your data stays in your account and the container makes no outbound calls.

Use is governed by the Causilo License v1.0, the end user license agreement for this listing: non-commercial research, testing and evaluation. Commercial or production use of Causilo or its outputs, or offering it as a hosted or API service, paid or free, needs a separate license from Nums AI Inc.; contact api@nums.world .