Prior labs unveils TabPFN-3.5, a tabular model that surpasses kaggle's 2015 champion without tuning

Prior Labs has launched TabPFN-3.5, a tabular foundation model that outperforms the 2015 Otto Kaggle competition's winning solution using default settings. The model achieves a log loss of 0.375 on the private leaderboard, surpassing the previous record of 0.382. Nick Erickson, co-creator of AutoGluon and AI researcher at Prior Labs, noted that TabPFN-3.5 ran on raw data without any tuning, achieving this result in about a minute on an RTX PRO 6000 GPU. TabPFN-3.5 also secures first place across seven benchmarks, including TabArena, BeyondArena, and STRABLE. The model was pretrained only on synthetic data and never exposed to Kaggle datasets. A reproducible Kaggle notebook is available for public review. Erickson highlighted the difficulty of improving from rank 50 to 10, which reduced log loss from 0.41 to 0.40, and further progress to the winning score required an additional 0.018. The model's base version scores 1866 Elo on TabArena, outperforming TabFM+ at 1823 and AutoGluon 1.6 extreme by 130 Elo in a fifth of the time. Despite having about 4x more parameters than TabPFN-3, the KV cache size remains similar, and single-row predictions match TabPFN-3's speed. However, on large training sets, the base model runs up to 2x slower than TabPFN-3.