TabPFN v2

TabPFN v2 is a transformer-based foundation model for tabular data from Prior Labs, published in Nature in January 2025 (Hollmann et al., 'Accurate predictions on small data with a tabular foundation model'). It uses prior-data fitted networks: a transformer pre-trained entirely on synthetic tabular datasets, so it can make predictions on a new small dataset in a single forward pass with in-context learning — no task-specific training, no hyperparameter tuning. It supports classification, regression, and generative tasks, handles up to 10,000 rows and 500 features, and was shown to beat tuned tree-based baselines on small-data benchmarks. The code is open source on GitHub (priorlabs/tabpfn) and it runs on CPU in seconds.

Limitations: the headline results apply to small-to-medium tabular data (up to ~10K rows); larger datasets need the newer 2.5/3.x line, whose weights are non-commercial. Licensing is nuanced: the TabPFN v2 code and weights sit under a modified Apache license with an attribution requirement (the code was relicensed to standard Apache 2.0 in v9.0.0, September 2026), while later model versions carry non-commercial weight licenses — so pin the v2 artifacts and read the license file if you deploy commercially. It is also not designed for non-tabular data or very large datasets.