TabPFN vs XGBoost in 2026: 1673 against 1375 Elo, and where trees still win

Summary. On the TabArena leaderboard, version 0.1.4, covering 51 curated datasets and retrieved on 10 June 2026, TabPFN-3 running its default configuration scores 1673 Elo. Tuned and post-hoc-ensembled XGBoost scores 1375. Tuned LightGBM, the strongest tree on the board, reaches 1433. For calibration, a default Random Forest is pinned at 1000 and a 400-point gap is roughly a 10 to 1 win rate. The foundation model gets there in one forward pass with no tuning, in 4.97 seconds of train time per 1,000 samples against LightGBM's 417. On the small-data slice, 36 of the 51 datasets, TabPFN-3 reaches 1642 and statistically ties AutoGluon's 4-hour ensemble at 1643. Prior Labs, the German lab behind TabPFN, is now part of SAP, and states a 93% win rate over classic machine learning as of July 2026. The peer-reviewed result that started this, published in Nature 637 in 2025, showed the same trick on datasets up to 10,000 rows in 2.8 seconds against ensembles tuned for 4 hours. For Indian teams scoring credit or churn on personal data, the DPDP Act 2023 attaches a ₹250 crore maximum penalty to getting the deployment wrong. The model decision is still not automatic, and the reasons are GPU dependence, row limits and benchmark protocol rather than the speed argument most summaries reach for.

That last point is where most coverage of this shift goes wrong, so it is worth stating plainly at the top. The common claim is that foundation models buy accuracy with latency. On the published board the default TabPFN-3 configuration is faster to fit than the tuned tree configurations it beats. What it actually costs you is a GPU, a ceiling on rows, and a dependency on a benchmark protocol that the benchmark's own authors call sensitive.

What the board says

TabArena is a living benchmark for tabular machine learning, published as arXiv:2506.16791 by Pichler, Salinas and colleagues. Its 51 datasets were curated from a pool of 1,053 candidates across 14 prior benchmarks, and every model runs under three protocols: defaults, tuned, and tuned with post-hoc ensembling. Results are aggregated as Elo with bootstrapped 95% confidence intervals. That design exists because the previous decade of tabular benchmarks contradicted each other, largely through dataset licensing problems, leakage and weak validation protocols.

Here is the overall board, best configuration per model family, as published on the CodeSOTA registry snapshot of the official TabArena data.

Read the third column before reacting to the second. AutoGluon at the top is not a model, it is a 4-hour multi-model ensemble listed as a reference ceiling. And the tree entries carry the cost of a full hyperparameter search, which is why LightGBM's 417 seconds per 1,000 samples dwarfs TabPFN-3's 4.97. CatBoost's default configuration is the fair like-for-like comparison against a default foundation model: it scores 1369 Elo at about 7 seconds per 1,000 samples.

So the honest one-line version is this. Against a default CatBoost, a default TabPFN-3 is both more accurate and slightly faster to fit, and needs a GPU to do it. Against a fully tuned LightGBM pipeline you already own and run on CPU, it is more accurate, and the migration question is about hardware and data shape rather than about accuracy.

The small-data slice is where the argument ended

36 of the 51 TabArena datasets are small, which is also the regime the original TabPFN work targeted and, more to the point, the regime most real business problems occupy. Churn tables, credit files, clinical cohorts and equipment logs are rarely millions of clean rows.

The top five single models in that slice are all foundation models. The best gradient-boosted tree trails the leader by about 253 Elo, which on this scale is roughly a 4 to 1 expected win rate. A single forward pass with no tuning statistically ties a 4-hour AutoML ensemble.

How TabPFN does it, and what the versions actually are

TabPFN is a transformer pre-trained on synthetic data that performs in-context learning: it does not train on your table, it conditions on it. The TabPFN v2 evaluation published in Nature 637, 319 to 326 by Hollmann and colleagues reported that a model pre-trained on around 100 million synthetic datasets outperformed every baseline on datasets up to 10,000 samples and 500 features, in 2.8 seconds, against ensembles tuned for four hours.

The version numbering has moved quickly enough that a lot of secondary coverage is stale, so it is worth pinning down from the vendor's own pages.

Prior Labs states that TabPFN-3 delivers a 93% win rate over classic machine learning on TabArena, predictions on 1 million samples in 0.2 seconds, and 20 times faster inference than previous versions, rising to 1,000 times faster at scale. Those figures come from the vendor, not from the independent board, and should be read as such. Dael Williamson, EMEA CTO at Databricks, describes the practical shape of it in a customer statement on the Prior Labs site: "Production-grade predictions in a single forward pass — typically measured in seconds." It is a vendor testimonial; the independent Elo numbers above are the ones to plan with.

Deployment has moved past the research-package stage. TabPFN is listed for Databricks, AWS SageMaker via the marketplace, and the Azure AI Foundry model catalog, with a Model Context Protocol endpoint for agent use. If you are building agent workflows that need a prediction rather than a generation, that MCP path is the interesting one, and it pairs with the retrieval decisions covered in our pgvector against dedicated vector database comparison.

The caveats the benchmark authors put in writing

The TabArena paper is unusually direct about the limits of its own leaderboard, and skipping this part is how teams end up with a migration that underdelivers.

Rankings are protocol-sensitive. Post-hoc ensembling of hyperparameter configurations reshuffles the order: without it, CatBoost beats the tuned neural networks that outrank it with it. The launch finding of the benchmark was that validation protocol and ensembling policy change rankings more than architecture choice does. If you run your own bake-off with a different harness, you will get a different order, and you will be measuring your harness.

Foundation-model dominance is a small-and-medium-data result. Context length bounds how much data fits into a forward pass, which is exactly why the version ceilings in the table above keep moving. On large tables, gradient-boosted trees hold the cost-efficiency line.

Hardware is not a footnote. Every foundation model in the top tier is marked GPU on the board. LightGBM, CatBoost and XGBoost run on CPU. For a team serving predictions from ordinary application servers, that difference is the entire operational argument, and it does not show up anywhere in an Elo column.

Interpretability has a measured price. The Explainable Boosting Machine, a glass-box additive model, sits at 1272 Elo, about 160 below LightGBM. If your model has to decompose into per-feature curves for a regulator, that is the cost of the property, and it is now a number rather than a debate.

A decision framework that survives a procurement conversation

Start with the shape of the table, not the leaderboard. Under roughly 50,000 rows, with a GPU available and no hard interpretability requirement, the foundation models are the accuracy frontier and require no tuning budget at all. Above that, or on CPU-only serving, LightGBM and CatBoost remain the sensible default, and XGBoost at 1375 Elo is still a solid model rather than a bad one.

Second, price the tuning budget honestly. A large part of the tree numbers on the board comes from tuned and post-hoc-ensembled runs. If your team does not actually run a hyperparameter search, your production XGBoost is not the 1375 configuration on that board. Comparing an untuned tree against a default foundation model is a fairer picture of most real pipelines.

Third, keep a tree in the loop during migration. Running both and comparing on your own holdout costs little, and it is the only evidence that transfers to your data. The benchmark tells you where to look; it does not tell you what your churn table does.

Fourth, decide the explainability question before the model question. Credit decisioning, insurance pricing and clinical triage in most jurisdictions need a defensible account of why a prediction came out the way it did. The 160 Elo you pay for a glass-box model may be cheaper than the alternative.

Fifth, watch what the change does to your feature pipeline. A large part of the practical appeal here is that TabPFN handles missing values, categoricals, outliers and mixed types on raw input, which removes preprocessing code rather than adding it. That is a maintenance saving, and it is also a loss of control: the feature engineering your team built up over years encodes domain knowledge that a general pre-trained model does not have. Prior Labs states that a tuned TabPFN outperforms an untuned one, so the preprocessing work does not become worthless, it becomes optional. Teams with genuinely informative engineered features should test with them retained rather than assume the model subsumes them.

Sixth, budget for evaluation as an ongoing cost rather than a one-off bake-off. Model versions in this space moved from v2 to 2.5 to 2.6 to 3 inside about eighteen months, and each release changed the size ceiling. Teams that wired a permanent evaluation harness are the ones that can take the next version without a project; the practices in our AI evaluation and observability work apply here directly.

India-specific considerations

The use cases driving tabular model adoption in India are the ones sitting on personal data: loan approvals, insurance underwriting, churn scoring for telecom and D2C, and clinical risk. That places the model choice inside the DPDP Act 2023, where the maximum penalty for failing to take reasonable security safeguards leading to a personal data breach is ₹250 crore.

Two consequences follow. First, deployment location matters as much as accuracy. A hosted prediction API means customer records leave your environment, so the on-premises and private-cloud deployment options are the ones to evaluate for regulated data, and the same reasoning we set out for running LLMs locally in production applies to tabular models. Second, automated decisions on individuals need an audit trail regardless of which model produced them; our DPDP engineering playbook for Indian startups covers the logging and consent architecture around that.

There is also a cost argument specific to smaller Indian teams. The tuning budget that a tuned-and-ensembled tree pipeline assumes is real engineering time. A default-configuration model that needs no hyperparameter search removes a recurring cost, which matters more at ten engineers than at a thousand. The offsetting cost is GPU access, which is the trade to price out before committing.

FAQ

How eCorpIT can help

eCorpIT builds and operates production prediction systems for teams working with churn, credit, demand and clinical data. Our senior engineering teams handle the parts that decide whether a benchmark result transfers: an evaluation harness on your own data, a fair comparison between tuned trees and foundation models, deployment that keeps regulated records inside your environment, and the monitoring that catches drift after launch. We are model-agnostic by design, because the right answer depends on your table shape and your serving constraints. Contact us for a scoped evaluation against your existing pipeline.

References

1. Tabular machine learning 2026: the TabArena leaderboard — CodeSOTA registry snapshot of TabArena v0.1.4 board data, retrieved 10 June 2026, with per-model Elo, confidence intervals and train and predict times.

1. Official TabArena leaderboard — the live board.

1. TabArena: a living benchmark for machine learning on tabular data — Pichler, Salinas et al., benchmark methodology and protocol-sensitivity findings.

1. Accurate predictions on small data with a tabular foundation model — Hollmann et al., Nature 637, 319 to 326.

1. TabPFN-3 model page — Prior Labs, current model claims, deployment options and customer statements.

1. TabPFN-2.5 model page — stated size ceilings and version history.

1. TabPFN-3 model report — Prior Labs technical report.

1. TabPFN-3 paper — the model paper linked from the TabArena board.

1. TabPFN-2.6 and RealTabPFN paper — successor line to the Nature model.

1. XGBoost: a scalable tree boosting system — Chen and Guestrin.

1. LightGBM — NeurIPS 2017.

1. CatBoost — the gradient boosting library used in the comparison.

1. AutoGluon — the AutoML reference pipeline at the top of the board.

1. TabArena benchmark code — open protocol if you want to reproduce the harness on your own data.

1. Penalties and adjudication under India's DPDP Act, 2023 — the ₹250 crore maximum for security-safeguard failures.

Last updated: 20 July 2026.