Show HN: TabPFN-2.5 – SOTA foundation model for tabular data (priorlabs.ai)

The TabPFN team announced TabPFN-2.5, a next‑generation tabular foundation model that scales to roughly 20× more data cells than TabPFNv2 (up to 50,000 rows and 2,000 features) and, on industry benchmarks, reportedly outperforms tuned tree models (e.g., XGBoost, CatBoost) while matching the accuracy of AutoGluon 1.4 — a four‑hour tuned ensemble that even included the prior TabPFNv2. Like its predecessors, TabPFN-2.5 is a meta‑trained, training‑free predictor: it’s pretrained on large synthetic distributions of tabular tasks and performs inference via in‑context learning (a forward pass) rather than dataset-specific gradient descent and hyperparameter tuning. It also continues to handle real‑world messiness (categorical features, missing values, outliers) and aims to give better calibrated uncertainty estimates out of the box. The release adds a pragmatic distillation engine that converts the large TabPFN-2.5 into compact MLPs or tree ensembles for production, preserving most of the model’s accuracy while delivering orders‑of‑magnitude lower latency and plug‑and‑play deployment. For the AI/ML community this means tabular foundation models are now viable for much larger, messier datasets and production settings, reducing the heavy tuning overhead of traditional approaches and broadening opportunities for transferable, data‑efficient tabular solutions.