TabPFN - A Foundation Model for Tabular Data

TabPFN is a foundation model dedicated to tabular data, allowing classification and regression to be performed immediately via a scikit-learn style interface using the fit / predict approach.

The base model, TabPFN-2.6, was trained on purely synthetic data, and automatically downloads a checkpoint on first run. No separate training pipeline is required, and raw data can be fed in directly without preprocessing such as scaling or one-hot encoding. Missing values are also handled internally.

On the operational side, GPU usage is recommended, requiring roughly 8GB VRAM or more. On CPU, only around 1,000 samples or fewer is practical, and for environments without a GPU, cloud inference via TabPFN Client is also provided.

There are also clear constraints on usage. Calling predict repeatedly on a per-sample basis recomputes the training set each time, which can be about 100x slower than a single call, so it is recommended to split the test set into batches of 1,000. The performance range centers on up to 100,000 samples and 2,000 features, with 50,000-100,000 samples requiring the ignore_pretraining_limits=True setting, and over 100,000 requiring the separate Large Datasets Guide.

A surrounding ecosystem is also provided.

• TabPFN Extensions: SHAP interpretation, outlier detection, synthetic data generation, embedding extraction, hyperparameter optimization, Post-Hoc ensembling
• HuggingFace checkpoints: for large features, for large samples, for small samples, and a version fine-tuned on real data
• Enterprise Edition: low-latency inference based on a distillation engine, supporting up to 10 million rows, with a commercial license
• TabPFN UX: a no-code graphical interface

The license is the Prior Labs License (Apache 2.0 + attribution requirement), and the TabPFN-2.5/2.6 weights are provided under a non-commercial license.