Introducing Scaling Mode for TabPFN: Foundation Models for Tabular Data with Strong Performance up to 10 Million Rows

Over the past two years, TabPFN has gone from a research prototype for tiny datasets to a production‑ready foundation model powering hundreds of tabular ML applications — from datasets with no more than 1,000 rows in 2023 to state-of-the-art performance for 100,000 rows three weeks ago.

Today, we’re excited to introduce Scaling Mode: a new operating mode that lifts the remaining constraints on dataset size. We now see strong performance with up to 10 million rows (and with no fixed upper limit; we just haven’t benchmarked beyond that).

Our new Scaling Mode builds directly on top of TabPFN-2.5. We demonstrate that we can leverage models pretrained on a smaller scale for data that is orders of magnitude larger. Scaling Mode enables TabPFN-2.5 to scale with the training data — the more data we use, the better the predictions.

The Accelerating Story of TabPFN

Our journey toward a million-row scale has accelerated rapidly:

• In May 2023, we introduced TabPFN-1 for tables with up to 1,000 rows.
• In January 2025, we scaled to tables with up to 10,000 rows with TabPFNv2.
• In November 2025, TabPFN-2.5 topped the TabArena leaderboard for tables with up to 100,000 rows.
• Today, less than 4 weeks after TabPFN-2.5, we introduce TabPFN-2.5 Scaling Mode, which supports an arbitrary number of rows, and which we benchmarked with strong performance up to 10 million rows. The limit on the number of features is 2,000, unchanged from TabPFN-2.5.

First Results

We share the first results of Scaling Mode on our internal benchmark below, containing internal datasets ranging from 1M to 10M rows, covering applications from industry and science. We compare to the three standard gradient boosting libraries CatBoost, XGBoost, LightGBM, and to TabPFN-2.5 (with a naive bootstrap approach that subsamples 50K different rows for each estimator).

The results show that Scaling Mode enables TabPFN-2.5 to continue improving performance as data volume increases: it scales dramatically better than TabPFN-2.5 with a naive subsampling approach, and there is no evidence of the gap to gradient boosting shrinking as we scale up.

What’s Next

For the time being, we are working with customers to apply TabPFN-2.5 Scaling Mode to their largest datasets.

If you are interested in this, please contact sales@priorlabs.ai

We are also preparing to make TabPFN-2.5 Scaling Mode more widely accessible to customers and are actively working on TabPFN-3, our next major model version, leading to even stronger performance on large datasets.