TabPFN-3-Plus

TabPFN-3-Plus by Prior Labs is the latest generation of the leading Tabular Foundation Model. Pretrained exclusively on synthetic data from a structural-causal-model prior, it produces calibrated predictions for classification and regression in a single forward pass with no model selection and no hyperparameter tuning. On the public TabArena benchmark, a forward pass tops every prior tabular foundation model and Pareto-dominates the speed-versus-performance frontier. Validated scale: up to 1M training rows at 200 features, 100k at 2000 features, or 1k at 20000 features. A redesigned architecture with per-row in-context learning, two-stage row compression, and row chunking runs up to 20x faster than TabPFN-2.5 and fits 1M rows on a single H100 GPU. Native many-class classification, mixed feature types, missing values, and outliers are handled directly. TabPFN-3-Plus also includes native text-feature support: string-valued columns (free-text fields such as product names, claim descriptions, or customer reviews) are accepted directly and encoded jointly with the numeric and categorical columns inside the model, so cross-feature interactions between text and structured columns are learned end-to-end rather than imposed by a fixed encoder. Released under the TABPFN-3.0 License v1.0 for research and internal evaluation; commercial licenses via Prior Labs.