Overview
TabPFN-2
The original foundation model for tabular data - published in Nature (2024). Proven academic benchmark for small-and medium-sized datasets.
TabPFN-TS
TabPFN adapted for time series forecasting - reframing temporal prediction as tabular regression for fast, zero-shot forecasts.
Our Flagship Model
TabPFN-2.5 is the latest generation of Prior Labs’ tabular foundation models - delivering state-of-the-art zero-shot performance across classification and regression tasks, with scalability up to 50,000 samples and no training required. It extends the original TabPFN v2 architecture with deeper layers, broader synthetic priors, and advanced preprocessing - providing the fastest and most accurate TabPFN model yet.- Zero-shot SOTA performance: Matches or surpasses tuned AutoGluon 1.4 ensembles in just 2.8 seconds, compared to 4 hours of tuning.
- Scales up to 50 k samples: Efficient inference on large tabular datasets using multi-GPU parallelism.
- 18-layer transformer - Deeper architecture improves generalization and calibration across diverse data types.
- Enhanced robustness - Combines robust scaling, soft clipping, quantile transforms, and SVD features for stability on outlier-heavy datasets.
- Meta-optimized by TabPFN - Its own hyperparameters were optimized using TabPFN v2 as a surrogate model, enabling faster, more efficient development through in-context optimization.
Commercial use restrictionsThe TabPFN-2.5 weights hosted on Hugging Face and GitHub are provided for research and non-commercial use only. For commercial licensing or large-scale deployment, please contact us.
Architecture
TabPFN-2.5 retains the alternating attention architecture introduced in TabPFN v2 - attending over both rows and features - while scaling to 18 transformer layers for improved expressivity and calibration. Key improvements include:- Broader synthetic training distribution: covers more diverse feature types and statistical structures.
- Optimized feature transformations: robust scaling and quantile mappings improve stability under noise and outliers.
- SVD-based augmentation: adds global structure awareness for high-dimensional datasets.
- Reduced memory footprint: removes redundant feature copies, cutting memory use nearly in half.