TabPFN-3 just released: a pre-trained tabular foundation model for up to 1M rows [R][N]

TabPFN-3 was released today, the next iteration of the tabular foundation model, originally published in Nature. Quick recap for anyone new to TabPFN: TabPFN predicts on tabular data in a single forward pass - no training, no hyperparameter search, no tuning. Built on TabPFN-2.5 (Nov 2025) and TabPFNv2 (Nature, Jan 2025), which together crossed 3M downloads and 200+ published applications. What's new: * Scale: 1M rows on a single H100 (10x larger than 2.5).A reduced KV cache (\~8GB per million rows per estimator) and row-chunked inference make this practical on a single GPU * Speed: 10x-1000x faster inference than previous versions. 120x on SHAP via KV caching * Thinking Mode (API only): test-time compute pushes predictions further via one-time extra fitting at inference. Beats every non-TabPFN method on TabArena by over 200 Elo, including 4-hour-tuned AutoGluon 1.5 extreme. Gap more than doubles to 420 Elo on the larger-data slice. * Accuracy: it has a 93% win rate over classical ML on TabArena * Many-class: native non-parametric retrieval decoder supporting up to 160 classes * Calibrated quantile regression: bar-distribution regression head produces calibrated quantile predictions in a single forward pass * Lifts adjacent tasks: time-series, interpretability, and new SOTA on relational benchmarks. * 3 deployment paths: API, enterprise licensing, and open-source weights (permissive for research and academic evaluation) You can try it [here](https://docs.priorlabs.ai/quic