TabPFN
AI Foundation model for tabular data
From Wikipedia, the free encyclopedia
TabPFN (Tabular Prior-data Fitted Network) is a machine learning model for tabular datasets proposed in 2022. It uses a transformer architecture.[1] It is intended for supervised classification and regression analysis on small- to medium-sized datasets, with TabPFN-2.5 supporting up to 50,000 rows and 2,000 features.[4]
| TabPFN | |
|---|---|
| Developers | Noah Hollmann, Samuel Müller, Lennart Purucker, Arjun Krishnakumar, Max Körfer, Shi Bin Hoo, Robin Tibor Schirrmeister, Frank Hutter, Leo Grinsztajn, Klemens Flöge, Oscar Key & Sauraj Gambhir [1] |
| Initial release | September 16, 2023[2][3] |
| Stable release | v2.5
/ November 6, 2025 |
| Written in | Python[3] |
| Operating system | Linux, macOS, Microsoft Windows[3] |
| Type | Machine learning |
| License | Apache License 2.0 |
| Website | github |
History
TabPFN was first introduced in a 2022 pre-print and presented at ICLR 2023.[2] TabPFN v2 was published in 2025 in Nature by Hollmann and co-authors.[1] The source code is published on GitHub under a modified Apache License and on PyPi.[5] Writing for ICLR blogs, McCarter states that the model has attracted attention due to its performance on small dataset benchmarks.[6] TabPFN v2.5, the next generation of the foundation model, was released on November 6, 2025.[4]
Prior Labs, founded in 2024, aims to commercialize TabPFN.[7]
Overview and pre-training
TabPFN supports classification, regression and generative tasks.[1] It leverages "Prior-Data Fitted Networks"[8] models to model tabular data.[1] By using a transformer pre-trained on synthetic tabular datasets,[2][6] TabPFN avoids benchmark contamination and costs of curating real-world data.[2]
TabPFN v2 was pre-trained on approximately 130 million such datasets.[1] Synthetic datasets are generated using causal models or Bayesian neural networks; this can include simulating missing values, imbalanced data, and noise.[1] Random inputs are passed through these models to generate outputs, with a bias towards simpler causal structures.[1] During pre-training, TabPFN predicts the masked target values of new data points given training data points and their known targets, effectively learning a generic learning algorithm that is executed by running a neural network forward pass.[1] The new dataset is then processed in a single forward pass without retraining.[2] The model's transformer encoder processes features and labels by alternating attention across rows and columns.[9] TabPFN v2 handles numerical and categorical features, missing values, and supports tasks like regression and synthetic data generation,[1] while TabPFN-2.5 scales this approach to datasets with up to 50,000 rows and 2,000 features.[4]
Since TabPFN is pre-trained, in contrast to other deep learning methods, it does not require costly hyperparameter optimization.[9]