Prior Labs Releases TabPFN-3.5, a Tabular Foundation Model That Beats Otto Kaggle Winner With Defaults
Prior Labs released TabPFN-3.5, a tabular foundation model that scores 0.375 on the 2015 Otto Kaggle private leaderboard, beating the winning 0.382 solution using raw data and default settings in about a minute, according to MarkTechPost. The company reports first place across seven tabular benchmarks, with open weights for non-commercial use and API or commercial license for production.
The Otto Group Product Classification Challenge ran on Kaggle in 2015 and drew 3,505 teams competing for $10,000. Entrants sorted products into nine categories using 93 obfuscated count features, and submissions were scored with multi-class log loss, where lower is better. The winning entry came from Gilberto Titericz and Stanislav Semenov, both of whom have held the world No. 1 Kaggle grandmaster ranking. Their solution was a multi-layer stack of 36 models built on hand-crafted features.
Nick Erickson, co-creator of AutoGluon and an AI researcher at Prior Labs, has chased that score for years. According to Erickson, AutoGluon placed rank 23 in its 2020 paper, rank 14 in 2023 with AutoGluon 1.0, and rank 9 in August 2026 with AutoGluon 1.6. He said moving from rank 50 to rank 10 cut log loss from 0.41 to 0.40, while reaching the winning 0.382 from rank 10 took a further 0.018, nearly double.
Erickson said TabPFN-3.5 ran on raw data with default settings and took about a minute on an RTX PRO 6000 GPU. The model was pretrained only on synthetic data and never saw Otto or any Kaggle dataset, and a reproducible Kaggle notebook is public.
The technical report lists first place on TabArena, BeyondArena, STRABLE, MulTaBench, RelArena-α, TALENT and ScoringBench. The top entry is not always the base model. TabPFN-3.5-Thinking holds first on TabArena, BeyondArena, STRABLE and MulTaBench, while an internal TabPFN-Rel harness preview takes RelArena-α.
On TabArena, a living benchmark of 51 datasets, Thinking reaches 1910 Elo. The base model scores 1866, ahead of TabFM+ at 1823. Prior Labs says the base model beats AutoGluon 1.6 extreme by 130 Elo in a fifth of the time.
BeyondArena spans 142 datasets with grouped, temporal, wide, text-rich and high-cardinality data. TabPFN-3.5 finishes about 150 Elo points ahead of the previous overall leader. Tuned and ensembled MLPs still lead on its grouped, temporal and large-data slices.
Under the hood, the in-context transformer grows from 512 to 1024 dimensions, and parameters rise to 220 million, up from 53 million for TabPFN-3 classification. A single multitask checkpoint now covers classification and regression. Values pass through learned Fourier features and in-context ECDF ranks, which do not change under monotonic transforms such as log scaling. Quantile transforms, robust scaling and SVD features are removed. The model supports up to 1 million rows, with 6,000 features recommended and 20,000 supported. Synthetic data now emphasizes high-cardinality, wide and grouped tables. The KV cache stays roughly the same size as TabPFN-3, despite about four times more parameters. Cached single-row predictions match TabPFN-3 speed, but on large training sets the base model runs up to twice as slow as TabPFN-3.
The model family includes TabPFN-3.5, with open weights, 220 million parameters and eight estimators by default; TabPFN-3.5-Fast (alpha), with open weights, 84 million parameters and up to six times the speed of the base model; TabPFN-3.5-Plus, available only through API and enterprise, adding native text handling and FP8 attention; and TabPFN-3.5-Thinking, which spends extra inference compute with no LLMs, real data or search and runs up to 12 times faster than TabPFN-3-Thinking.
Open weights run locally for research, evaluation and Kaggle, but production use requires Prior Labs' API or a commercial license. MarkTechPost reported the release and linked the technical report and technical details.