Prior Labs Releases TabPFN-3.5: A Tabular Foundation Model That Beats the Winning Otto Kaggle Solution With Default Settings

Prior Labs has put out TabPFN-3.5, the latest version of its tabular foundation model, and the headline is hard to miss: it beat the winning solution from Kaggle’s 2015 Otto Group Product Classification Challenge using raw data and default settings. According to Nick Erickson, the run scored 0.375 on the private leaderboard, ahead of the long-standing winning score of 0.382. The model took about a minute on an RTX PRO 6000 GPU.

That Otto result still matters because it was never a toy problem. The contest drew 3,505 teams, offered $10,000, and asked people to sort products into nine categories from 93 obfuscated count features. The winning entry came from Gilberto Titericz and Stanislav Semenov, both world #1 Kaggle grandmasters, and it used a multi-layer stack of 36 models with hand-crafted features. Prior Labs says its model had never seen Otto or any Kaggle dataset at all; it was pretrained only on synthetic data.

The broader pitch is that TabPFN-3.5 is not just a one-off stunt. Prior Labs says the technical report puts it in first place on seven tabular benchmarks: TabArena, BeyondArena, STRABLE, MulTaBench, RelArena-α, TALENT and ScoringBench. The base model is not always the top entry in those comparisons, though. On some benchmarks, the company’s Thinking variant or an internal Rel harness preview takes the lead instead.

Under the hood, the model got bigger and simpler at the same time. The in-context transformer grows from 512 to 1024 dimensions, parameters rise to 220M from 53M in TabPFN-3 classification, and a single checkpoint now handles both classification and regression. The preprocessing stack also got stripped down: quantile transforms, robust scaling and SVD features are gone, replaced by learned Fourier features and in-context ECDF ranks that stay stable under monotonic transforms like log scaling.

Prior Labs is also making a sharp distinction between what is open and what is commercial. The open weights can run locally for research, evaluation and Kaggle, but production use needs Prior Labs’ API or a commercial license. There are also separate variants in the family: a smaller TabPFN-3.5-Fast alpha, a Plus version with native text handling and FP8 attention for API and enterprise customers, and a Thinking model that spends extra inference compute without using LLMs, real data or search.