¶ TabPFN
TabPFN is a transformer-based foundation model for tabular data prediction, designed to solve classification and regression tasks in a single forward pass using in-context learning. It is described in the original paper TabPFN: A transformer that solves small tabular classification problems in a second and in the newer report TabPFN-2.5: Advancing the State of the Art in Tabular Foundation Models.
Code, installation, examples, and the wider ecosystem are available in the official GitHub repository.
¶ Model description
TabPFN is a foundation model specialized for structured tabular datasets. Instead of training a separate model from scratch for every dataset, it uses a pretrained transformer that performs prediction directly through in-context learning, making it especially effective on small and medium-sized tabular problems.
The newer TabPFN-2.5 release extends the scope of the model substantially. It is designed for datasets with up to 50,000 rows and 2,000 features, which is a major increase over earlier TabPFN versions. The model supports both classification and regression, and the ecosystem around it adds capabilities such as interpretability, embeddings, unsupervised tasks, ensembling, and hyperparameter optimization.
Key traits of TabPFN:
- Tabular foundation model: Built specifically for structured tabular prediction tasks.
- In-context learning: Solves prediction tasks in a forward pass without conventional dataset-specific training from scratch.
- Supports classification and regression: Separate interfaces are available for both task types.
- Strong small-to-medium data performance: Particularly suited for datasets where classic tabular ML is usually dominant.
- Extended ecosystem: Includes tools for interpretability, embeddings, outlier detection, ensembling, and more.
¶ TabPFN
Figure 1 (from the official release materials) illustrates the workflow and positioning of TabPFN:
- A tabular dataset is given as input, with rows as samples and columns as structured features.
- The pretrained transformer-based foundation model performs prediction directly through in-context learning.
- Separate estimator interfaces are provided for classification and regression tasks.
- The broader TabPFN ecosystem extends the core model with interpretability tools, unsupervised utilities, embeddings, hybrid methods, and post-hoc ensembles.
- For practical use, GPU inference is recommended, although small datasets can also be handled on CPU.
¶ Intended uses & limitations
TabPFN is intended for:
- Tabular classification and regression on structured datasets.
- Small and medium-sized datasets where strong accuracy is needed with minimal tuning effort.
- Rapid experimentation and prototyping in tabular machine learning workflows.
- Research and analysis workflows involving feature importance, embeddings, outlier detection, and tabular model comparison.
Limitations:
- TabPFN is intended for structured tabular data, not for unstructured modalities such as raw text, images, or audio.
- Practical performance degrades outside the intended regime, especially beyond roughly 50,000 rows or 2,000 features.
- GPU use is strongly recommended for efficient inference, while CPU use is only practical for relatively small datasets.
- The default TabPFN-2.5 model weights are released under a non-commercial license, so production or commercial use requires special attention to licensing terms.
¶ BibTeX entry and citation info
@misc{grinsztajn2025tabpfn,
title={TabPFN-2.5: Advancing the State of the Art in Tabular Foundation Models},
author={Léo Grinsztajn and Klemens Flöge and Oscar Key and Felix Birkel and Philipp Jund and Brendan Roof and
Benjamin Jäger and Dominik Safaric and Simone Alessi and Adrian Hayler and Mihir Manium and Rosen Yu and
Felix Jablonski and Shi Bin Hoo and Anurag Garg and Jake Robertson and Magnus Bühler and Vladyslav Moroshan and
Lennart Purucker and Clara Cornu and Lilly Charlotte Wehrhahn and Alessandro Bonetto and
Bernhard Schölkopf and Sauraj Gambhir and Noah Hollmann and Frank Hutter},
year={2025},
eprint={2511.08667},
archivePrefix={arXiv},
url={https://arxiv.org/abs/2511.08667},
}