Prior Labs just dropped TabPFN-3.5, and it’s doing something that used to take teams months of feature engineering in about sixty seconds. The new model scored 0.375 on the private leaderboard of the 2015 Otto Group Product Classification Challenge, beating the winning score of 0.382 set by Kaggle grandmasters Gilberto Titericz and Stanislav Semenov. That original win required a complex stack of 36 models and hand-crafted features. TabPFN-3.5 achieved its result using raw data, default settings, and a single forward pass on an RTX PRO 6000 GPU. It never saw the Otto dataset during training.
This isn’t just a party trick for old competitions. The model claims first place across seven major tabular benchmarks, including TabArena and BeyondArena. On TabArena, the “Thinking” variant hits 1910 Elo, leaving competitors like TabFM+ behind. Prior Labs expanded the architecture significantly, jumping from 53 million parameters in TabPFN-3 to 220 million here. They also unified classification and regression into a single checkpoint. New encoding methods, specifically learned Fourier features and in-context ECDF ranks, help the model handle monotonic transforms without breaking a sweat. Preprocessing steps like quantile transforms and SVD are gone, simplifying the pipeline drastically.
However, there is a catch for enterprise users. While the weights are open for research and Kaggle competitions, production use requires a commercial license or access to Prior Labs’ API. The base model supports up to 1 million rows and 20,000 features, but it runs up to twice as slow as its predecessor on large training sets. For speed demons, the alpha release of TabPFN-3.5-Fast offers a smaller 84 million parameter model that runs up to six times faster. This shift suggests tabular AI is moving away from manual tuning toward foundation models that treat tables like language, though the compute cost for inference remains a consideration for large-scale deployment.
