Tabular Foundation Model from Prior Labs for classification, regression, and forecasting on tabular data in a single forward pass.
Context window
50K rows × 2K features
tokens
Parameters
Tabular foundation model
parameters
Access:APIDownloadHostedDeployment:💻 Local☁ Cloud
Overview
Applications
Access & deployment
APIDownloadHosted
LocalCloud
Weights: Open weights
Key parameters
📏 Context: 50K rows × 2K features
🧩 Parameters: Tabular foundation model
✓ Fine-tuning
📥 Input: structured data, time series
Technical specification
Context window
50K rows × 2K features
tokens
Parameters
Tabular foundation model
parameters
License
Apache-2.0 + non-commercial license dla wag (TabPFNv2/2.5); komercyjne wykorzystanie wymaga API/Enterprise
Hardware requirements
GPU recommended for inference (CUDA); CPU mode available for smaller datasets. Runs on a single consumer GPU for datasets <10K rows.
Features:✓ Fine-tuning
Modalities
⬇ Input
structured_datatime_series
⬆ Output
structured_data
Capabilities and applications
Native model capabilities
Classification
Assigning an observation to one of predefined classes (binary or multi-class). Output: class label and optionally probabilities.
Category: other
Regression
Predicting a continuous numerical value (e.g., price, temperature, risk) based on input features.
Category: other
Time-series forecasting
Predicting future values of a time series from historical observations, seasonality, and external regressors.
Category: other
Anomaly detection
Identifying observations that deviate from the normal pattern — key in fraud detection, machine monitoring, and security.
Category: other
Synthetic data generation
Generating synthetic datasets that preserve the statistical properties of the original — used for model training, testing, and privacy protection.
Category: structured_generation
Tabular prediction
Prediction on tabular data (rows × columns) — classification, regression, or time series — the domain of tabular foundation models like TabPFN.
Category: other
Interpretability
Explainability of model predictions: identifying features driving the decision (SHAP, feature importance, attention weights, counterfactuals).
Category: other
Benchmark results
1 benchmark
TabArena
accuracy · zero-shot, dataset size up to 50K samples, 2K features
matches AutoGluon 1.4 (4h tuned)
📄 TabPFN-2.5 Model Report (Prior Labs)
According to the technical report, TabPFN-2.5 outperforms all tuned tree-based models and matches AutoGluon 1.4 (an ensemble that itself includes TabPFN v2 tuned for 4 hours).