Prior Labs has launched TabPFN-3.5, a tabular foundation model that outperformed the winning solution of the 2015 Otto Group Product Classification Challenge on Kaggle. The model achieved a log loss score of 0.375 on the private leaderboard using default settings and raw data, surpassing the previous winning score of 0.382.
The model was trained exclusively on synthetic data and did not utilize any Kaggle datasets during pretraining. It completed the task in approximately one minute on an RTX PRO 6000 GPU. While the base model is available with open weights for research and evaluation, production deployment requires a commercial license or the use of the Prior Labs API.
TabPFN-3.5 leads several benchmarks, including TabArena, BeyondArena, STRABLE, and MulTaBench. On the TabArena benchmark, which covers 51 datasets, the model reached 1866 Elo, outperforming AutoGluon 1.6. A variant, TabPFN-3.5-Thinking, achieved 1910 Elo on the same benchmark.
The architecture of TabPFN-3.5 features approximately four times more parameters than its predecessor, TabPFN-3, while maintaining a similar KV cache size. Despite this, the model operates up to two times slower than TabPFN-3 on large training sets, though single-row prediction speeds remain consistent.
