Prior Labs releases TabPFN-3.5 for tabular prediction
Prior Labs has released TabPFN-3.5, a tabular foundation model that makes predictions in a forward pass without training or tuning for each dataset. The company reports first-place results on seven tabular benchmarks. Its separate Otto demonstration produced a private-leaderboard log loss of 0.375, compared with 0.382 for the winning 2015 Kaggle entry by Gilberto Titericz and Stanislav Semenov. TabPFN-3.5 used the raw data and default settings, running in about one minute on an RTX PRO 6000 GPU. The model had not been trained on Otto or other Kaggle datasets.
The Otto competition classified products into nine categories from 93 obfuscated count features. The former winning solution combined 36 models with hand-crafted features. On TabArena, TabPFN-3.5-Thinking reaches 1910 Elo across 51 datasets, while the base model scores 1866. Prior Labs says the model also leads BeyondArena overall, although tuned and ensembled MLPs remain ahead on some data slices.
The new base model has 220M parameters, up from 53M in TabPFN-3, and uses learned Fourier features and in-context ECDF ranks. Open weights are available for research, evaluation and Kaggle, but commercial production use requires the API or a license.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.