TabPFN-3.5: no-training tabular ML beats a Kaggle-winning stack
Prior Labs' open foundation model handles 1M rows and 20K features, tops seven tabular benchmarks, and installs in one pip command. Production needs a license.
Copy markdown
One checkpoint, seven benchmark wins
TabPFN-3.5 predicts both classification and regression from a single tabular foundation model — no feature engineering, no per-dataset training — and Prior Labs reports first place across seven benchmarks. Point it at a raw table and get calibrated predictions in seconds.
It beat a Kaggle-winning 36-model stack on raw data
On the classic Otto Kaggle set it scored 0.375 log loss, edging the 2015 winner's 0.382 — a solution that stacked 36 hand-crafted models. TabPFN ran on raw data with default settings in about a minute on one GPU.
Now it scales to messy, real tables
The ceiling jumps to 1M rows and 20,000 features (6,000 recommended), with native handling of text, high-cardinality categories, and wide tables — the enterprise data that broke earlier TabPFN. Skewed targets like claim amounts get proper predictive distributions.
Where it lands on TabArena
The base model hits 1866 Elo on TabArena versus 1823 for TabFM+, and beats AutoGluon 1.6 extreme by 130 Elo in a fifth of the time. A Plus variant tops string-heavy tables, a Thinking variant trades speed for accuracy, and a Fast alpha runs up to 6x quicker.
How to run it today
pip install tabpfn pulls the open weights — free for research, evaluation, and Kaggle under a non-commercial license; production needs the hosted API or a commercial license. It's also on REST, AWS SageMaker, Microsoft Foundry, and SAP AI Core, with API charges cut 50% through Sept 29.
Build this weekend: a no-train predictor for CSVs
Drop a churn, lead-scoring, or claim-amount table in and get a strong first-pass model without touching XGBoost or hyperparameters — often beating a hand-tuned baseline out of the box. A same-day way to sanity-check whether an ML feature is even worth building.