Prior Labs releases TabPFN-3.5, beating 2015 Otto Kaggle winner with default settings

Prior Labs has released TabPFN-3.5, a tabular foundation model that predicts on a table in a single forward pass without per-dataset training or tuning. The model scores 0.375 on the private leaderboard of the 2015 Otto Group Product Classification Challenge, beating the original winning score of 0.382 achieved by Gilberto Titericz and Stanislav Semenov.

• TabPFN-3.5 achieves first place across seven tabular benchmarks: TabArena, BeyondArena, STRABLE, MulTaBench, RelArena-α, TALENT, and ScoringBench.
• The base model grows to 220M parameters with a wider in-context transformer (1024 dimensions) and uses learned Fourier features and ECDF ranks for cell encodings.
• A single multitask checkpoint now covers both classification and regression, removing the need for quantile transforms, robust scaling, and SVD features.
• The model supports up to 1M rows with 6,000 recommended features (20,000 supported) and runs in about a minute on an RTX PRO 6000 GPU.
• The family includes TabPFN-3.5-Fast (84M parameters), TabPFN-3.5-Plus (API/enterprise with text handling), and TabPFN-3.5-Thinking (extra inference compute).

The release provides open weights for research, evaluation, and Kaggle, while production use requires Prior Labs' API or a commercial license.