Skip to content
Contact UsStart Building

Causilo / Tabular Foundation Model

Different questions, one Causilo

Predict churn, equipment condition or product demand. Your examples define the task. Causilo brings pretrained classification and regression.
QuickstartExplore applications
Causilo
Pretrained model

Who is likely to leave?

Customer history
AccountUsage / mo.TenureChurn
012642 visits18 mo.Stayed
03913 visits6 mo.Left
065828 visits24 mo.Stayed
08424 visits8 mo.?
Predicted churn riskHigh riskAccount 0842
Your next step

Prioritize this account for outreach.

Illustrative data and predictions

More features, less overhead

Feature-side attention stays linear as the number of features grows. Compare measured feature-mixing runtime with five other foundation models.
1.23
Causilo growth exponent

Feature-mixing modules

1.90–1.99
Other evaluated TFMs

Fitted over 16–65,536 features

Relative runtime · log scale

Number of features · log scale

Causilo · 65,536 features · ≈ 12.05× its 16-feature runtime

Feature-mixing modules · Normalized per model

Benchmark details

Selected feature-mixing operations with 16 fixed rows. Each model is normalized to its own runtime at 16 features; these curves compare growth, not absolute latency.

20 warm-up runs and 50 CUDA-event timings per point. Lines connect measured medians; the reported exponent fits T(F) = c + aFᵝ. This is not an end-to-end inference benchmark.

Plot coordinates reconstructed from vector figures. Approximate values; headline values come from the report text.

Technical report · Figure 3

Accuracy and speed, on the frontier

Compare the full field. Causilo sits at the fast end of the high-accuracy frontier in the reported TabArena evaluation.
1,785
Combined TabArena Elo
0.104 s
Median inference / 1K test rows

TabArena Elo ↑

Inference / 1K test rows · log scale

Causilo · ≈ 1,785 Elo · ≈ 0.104 s / 1K test rows

TabArena · All configurations in report Figure 1

Benchmark details

September 18, 2026 snapshot · 51 datasets. Higher Elo and lower inference time are better. The dashed line is the empirical accuracy–speed frontier among the displayed configurations.

Figure 1 includes foundation models, conventional models and ensemble/system configurations. Circles and diamonds preserve those figure markers. Hardware and configuration differ; see the report for the evaluation protocol.

Plot coordinates reconstructed from vector figures. Approximate values; headline values come from the report text.

Technical report · Figure 1

Try it on your own data

Start with a held-out part of your table and compare predictions with its known outcomes.
Use the released Python package with your context rows, known targets and new rows.
X_context and y_context contain known examples. X_query contains held-out rows without their target values.
Open WeightPython
# pip install causilo
from causilo import CausiloRegressor

model = CausiloRegressor()
# Prepare known examples without updating model weights
model.fit(X_context, y_context)
predictions = model.predict(X_query)
Quickstart