Open tabular models now beat tuned XGBoost below 50k rows, no tuning

A no-training benchmark went 14/14 on AUC vs tuned XGBoost; TabICLv2 is pip-installable and TabPFN-2.5 handles 50k rows — but wide tables still break it.

Nowline SEP 28 6:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • The 14/14 benchmark builders are passing around

    An independent RTX 4070 Ti run on 14 datasets (3,000 rows each, from Grinsztajn's suite) hit Hacker News today: TabICL took 14/14 on AUC and 12 wins, 1 tie, 1 loss on accuracy against a tuned XGBoost 3.4.1. Inference was about 0.8s per dataset versus 0.7–27s of XGBoost tuning — the "model that doesn't train" won without a hyperparameter search.

  • TabICLv2 is the open one you can pip install

    SODA-Inria's TabICLv2 (`pip install tabicl`, open weights) is state-of-the-art on the TabArena and TALENT benchmarks and still beats heavily-tuned XGBoost, CatBoost, and LightGBM on ~80% of datasets with zero tuning. On an H100 it fits and predicts 50k rows × 100 features in under 10 seconds — a real drop-in for small-to-mid tabular classification and regression today.

  • TabPFN-2.5 pushes the ceiling to 50k rows, 2k features

    Prior Labs' TabPFN-2.5 handles 5× the rows and 4× the features of v2 and matches AutoGluon 1.4 tuned for four hours — in a single forward pass. A new distillation engine compiles it into a compact MLP or tree ensemble, so you get most of the accuracy at low latency for plug-and-play deployment.

  • The caveats before you rip out XGBoost

    The edge is real below ~50k rows and on narrow tables, and it collapses outside that. TabPFN scored 0.7567 on a 419-column table — worse than untuned XGBoost — independent tests crashed past 230k rows, and GPU inference can run ~$27 versus cents for LightGBM on CPU. Inventory your models by row count and keep a tuned LightGBM as the control.