Prior Labs Releases TabPFN-3.5, Reporting a Win Over Otto’s 2015 Kaggle Score

Prior Labs has released TabPFN-3.5, a model for making predictions from spreadsheet-like data without training or tuning it separately for each dataset. Its headline result is a reported 0.375 score on Kaggle’s Otto challenge, below the 0.382 score of the competition’s 2015 winner—but the practical choice is not simply between a new model and an old leaderboard.

Tabular data is the structured information found in rows and columns: a product catalog, customer list, or business dataset. Prior Labs says TabPFN-3.5 can process such a table in a forward pass, meaning it produces predictions directly rather than being fitted anew to that specific dataset. The company’s pitch is a simpler starting point for problems that have traditionally demanded preprocessing and model selection.

A default run against a hand-built stack

The Otto Group Product Classification Challenge asked teams to sort products into nine categories using 93 obfuscated count features. Its private leaderboard used multi-class log loss, where lower scores are better. The 2015 winning entry combined 36 models with hand-crafted features; Prior Labs says TabPFN-3.5 beat that score on raw data, using default settings, in about one minute on an RTX PRO 6000 GPU.

That comparison makes the release notable, but it does not establish that default settings will beat tailored systems on every table. On BeyondArena, Prior Labs says TabPFN-3.5 was about 150 Elo points ahead of the previous overall leader, while tuned and ensembled multilayer perceptrons still led on grouped, temporal, and large-data slices.

One family, different tradeoffs

The base model has 220 million parameters, up from 53 million in TabPFN-3, and uses a 1,024-dimensional in-context transformer. Its cell encoding combines learned Fourier features with empirical cumulative distribution function ranks, a representation that does not change under monotonic transformations such as log scaling. Prior Labs also says the model supports tables with up to one million rows, recommends up to 6,000 features, and supports up to 20,000.

The release separates speed, capability, and access

• The base TabPFN-3.5 model has open weights for research, evaluation, and Kaggle use, but production requires Prior Labs’ API or a commercial license.
• TabPFN-3.5-Thinking spends additional inference compute and, according to Prior Labs, leads on TabArena, BeyondArena, STRABLE, and MulTaBench.
• The release’s seven reported first-place benchmark results apply to the TabPFN-3.5 family, not necessarily to the base model alone.

The result to test is the workflow

The central promise is not merely a better benchmark score. It is replacing much of the work of adapting a prediction system to each new table with a general model that can begin from defaults. The central constraint is equally concrete: teams seeking to put it into production must use Prior Labs’ API or obtain a commercial license, while the company’s top benchmark positions sometimes rely on a variant that uses extra inference compute.