A new independent benchmark says you can skip training a model entirely and still beat XGBoost on tabular data.
The comparison, published on a personal blog and surfaced on Hacker News, tests two tabular foundation models, TabPFN and TabICL, against XGBoost, the gradient-boosting library most practitioners reach for on structured data. TabPFN and TabICL skip the training step entirely: instead of fitting a model to your dataset, they take the data as direct input and produce predictions immediately. The author tuned the XGBoost baseline rather than leaving it at default settings, a deliberate choice to make the fight harder. Across all 14 datasets in the test, both no-training models came out ahead of that tuned XGBoost.
The tuned baseline is the detail that makes this notable. Plenty of foundation-model benchmarks only ever beat a lazy, default-settings XGBoost, which is an easy target and an easy result to dismiss. Beating a properly tuned one is a harder bar to clear, and it raises a real question for teams that currently budget engineering time on hyperparameter search and retraining pipelines.
A 14-0 sweep against a tuned baseline is a real result, not a rigged demo, and it's a fair reason to test TabPFN and TabICL against your own tabular data before reaching for XGBoost by default.