TabPFN

TabPFN is a tabular foundation model developed by Prior Labs that delivers state-of-the-art predictions on structured/tabular data for classification and regression tasks. It is trained purely on synthetic data and leverages a transformer architecture to make accurate predictions on datasets with up to 100,000 rows and 2,000 features. TabPFN supports local GPU inference as well as cloud-based inference via the TabPFN Client, and integrates with a rich ecosystem of extensions for interpretability, hyperparameter optimization, ensembling, and more.

• TabPFNClassifier & TabPFNRegressor: Install via pip install tabpfn, then call .fit() and .predict() with a scikit-learn-compatible API — no data preprocessing required.
• Multiple model versions: Choose between TabPFN-2.5 and TabPFN-2.6 checkpoints, including variants specialized for large features, large samples, or real-data fine-tuning.
• Cloud inference via TabPFN Client: Use the hosted API client for GPU-free inference — ideal for environments without local GPU resources.
• TabPFN Extensions ecosystem: Access advanced utilities including SHAP-based interpretability, outlier detection, synthetic data generation, many-class classification, hybrid Random Forest approaches, and automated HPO.
• Missing value handling: TabPFN natively handles missing values without requiring imputation preprocessing.
• KV Cache for fast prediction: Enable fit_mode='fit_with_cache' to speed up repeated predictions at the cost of additional memory.
• Offline/headless support: Download model weights manually from HuggingFace and configure via environment variables for CI or air-gapped environments.
• Enterprise Edition: Contact Prior Labs for a commercial license with Fast Inference Mode (distillation to MLP/tree ensemble), Large Data Mode (up to 10M rows), and dedicated support.
• No-code UI: Use TabPFN UX at ux.priorlabs.ai for a graphical interface to explore TabPFN capabilities without writing code.
• Python 3.9–3.13 support: Compatible with modern Python versions; GPU (8GB+ VRAM) recommended for optimal performance on larger datasets.