Prior Labs TabPFN-3.5 Beats 2015 Kaggle Otto Winner

Prior Labs has released TabPFN-3.5, an updated version of its tabular foundation model designed for efficient prediction on tabular data without per-dataset training or tuning. The model achieved first place across seven tabular benchmarks, according to Prior Labs. In a notable demonstration, TabPFN-3.5 surpassed the winning solution of the 2015 Otto Group Product Classification Challenge on Kaggle, a competition that attracted 3,505 teams vying for a $10,000 prize. Participants in the 2015 challenge were tasked with categorizing products into nine classes using 93 obfuscated count features, with submissions scored using multi-class log loss, where a lower score indicated better performance. The original winning entry was developed by Gilberto Titericz and Stanislav Semenov, both former Kaggle grandmasters, and comprised a multi-layer stack of 36 models built upon hand-crafted features. Nick Erickson, a co-creator of AutoGluon and an AI researcher at Prior Labs, has reportedly been pursuing the Otto score for years. Erickson stated that AutoGluon achieved rank 23 in a 2020 paper, improved to rank 14 with AutoGluon 1.0 in 2023, and reached rank 9 with AutoGluon 1.6 in August 2026. He noted that the progression from rank 50 to rank 10 involved reducing log loss from 0.41 to 0.40, and reaching the winning score of 0.382 from rank 10 required an additional 0.018 reduction, nearly doubling the effort. TabPFN-3.5 achieved a score of 0.375 on the private leaderboard for the Otto challenge. Erickson indicated that the model operated on raw data with default settings, completing the task in approximately one minute on an RTX PRO 6000 GPU. Importantly, the model was pretrained solely on synthetic data and had no prior exposure to the Otto dataset or any other Kaggle datasets. A reproducible Kaggle notebook detailing this achievement has been made publicly available. Beyond the Otto challenge, the technical report for TabPFN-3.5 lists first-place finishes on several other benchmarks, including TabArena, BeyondArena, STRABLE, MulTaBench, RelArena-α, TALENT, and ScoringBench. In some instances, a variant named TabPFN-3.5-Thinking secured first place on TabArena, BeyondArena, STRABLE, and MulTaBench. The deployment of TabPFN-3.5 is available through a license. While open weights are provided for local research, evaluation, and Kaggle competitions, production use necessitates Prior Labs’ API or a commercial license.