# TabPFN

> TabPFN is an open-source tabular foundation model that performs accurate classification and regression on small-to-medium datasets in seconds, trained purely on synthetic data using PyTorch.

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.*

## Features
- Tabular classification and regression
- Scikit-learn compatible API
- No data preprocessing required
- Native missing value handling
- GPU and CPU inference
- Cloud-based inference via TabPFN Client
- Multiple model checkpoints (v2.5, v2.6)
- SHAP-based interpretability
- Outlier detection and synthetic data generation
- Automated hyperparameter optimization (HPO)
- Post-hoc ensembling
- Many-class classification support
- Time-series feature support
- KV Cache for fast repeated predictions
- Offline/headless model usage
- Enterprise Edition with Fast Inference and Large Data Mode

## Integrations
PyTorch, CUDA, scikit-learn, SHAP, SHAP IQ, HuggingFace, Google Colab, Random Forest, uv (package manager)

## Platforms
WINDOWS, MACOS, LINUX, API, DEVELOPER_SDK, CLI

## Pricing
Open Source, Free tier available

## Version
2.6

## Links
- Website: http://priorlabs.ai
- Documentation: https://priorlabs.ai/docs
- Repository: https://github.com/PriorLabs/TabPFN
- EveryDev.ai: https://www.everydev.ai/tools/tabpfn