Prior Labs’ TabPFN-3.5: A Breakthrough Tabular Foundation Model Outperforms Kaggle Champions

A New Era for Tabular Data: Introducing TabPFN-3.5

At a glance, The world of artificial intelligence continues to push boundaries, and the latest innovation from Prior Labs marks a significant leap forward for tabular data. Their new foundation model, TabPFN-3.5, has not only demonstrated state-of-the-art performance across numerous benchmarks but has also achieved a remarkable feat: surpassing the winning solution of a prestigious 2015 Kaggle competition, all without any dataset-specific training or tuning.

Meanwhile, This release signals a paradigm shift in how we approach machine learning with structured data, offering a powerful, ready-to-use model that can predict on tables in a single forward pass.

The Otto Kaggle Challenge: A Historic Feat Revisited

One of the most compelling demonstrations of TabPFN-3.5’s capabilities comes from its performance on the 2015 Otto Group Product Classification Challenge. This Kaggle competition, which attracted over 3,500 teams, tasked participants with classifying products into nine categories based on 93 obfuscated features. The winning solution, crafted by Kaggle Grandmasters Gilberto Titericz and Stanislav Semenov, was an intricate stack of 36 models built upon meticulously hand-crafted features, achieving a multi-class log loss of 0.382.

In practical terms, Years later, Prior Labs’ AI researcher Nick Erickson, a co-creator of AutoGluon, finally saw TabPFN-3.5 achieve what many thought impossible. Running on raw data with its default settings, TabPFN-3.5 recorded an impressive log loss of 0.375 on the private leaderboard. This accomplishment was achieved in approximately one minute on an RTX PRO 6000 GPU, and critically, the model was pretrained solely on synthetic data, having never encountered the Otto dataset or any other Kaggle data.

Dominating Across Diverse Benchmarks

Beyond its impressive Kaggle victory, TabPFN-3.5 has established itself as a frontrunner across a spectrum of tabular benchmarks. The technical report from Prior Labs highlights its first-place ranking on seven distinct benchmarks, including TabArena, BeyondArena, STRABLE, MulTaBench, RelArena-α, TALENT, and ScoringBench.

• TabArena: On this living benchmark of 51 datasets, TabPFN-3.5-Thinking achieved an Elo score of 1910, with the base model scoring 1866—significantly ahead of competitors like TabFM+.
• BeyondArena: Spanning 142 datasets with varied characteristics (grouped, temporal, wide, text-rich, high-cardinality), TabPFN-3.5 leads the previous overall leader by about 150 Elo points.

For example, This widespread success underscores the model’s robustness and adaptability across diverse tabular data challenges.

Under the Hood: Innovations in TabPFN-3.5

Prior Labs implemented several key technical enhancements to achieve this new level of performance in TabPFN-3.5:

• Wider Model: The in-context transformer’s dimensions have expanded from 512 to 1024, increasing the parameter count to 220 million (up from 53 million in TabPFN-3).
• Unified Checkpoint: A single multitask checkpoint now efficiently handles both classification and regression tasks.
• Enhanced Cell Encodings: Values are processed through learned Fourier features and in-context ECDF ranks, which remain consistent under monotonic transformations.
• Streamlined Preprocessing: Complex steps like quantile transforms, robust scaling, and SVD features have been removed, simplifying the pipeline.
• Increased Scale: The model now supports up to 1 million rows, with recommendations for 6,000 features and support for up to 20,000.
• Refined Prior: The synthetic data used for pretraining now places greater emphasis on high-cardinality, wide, and grouped tables.

That said, Despite the substantial increase in parameters, the KV cache size remains comparable to TabPFN-3, allowing for similar speeds in cached single-row predictions, though training on large datasets may be up to twice as slow.

Accessibility and the TabPFN Ecosystem

Prior Labs offers TabPFN-3.5 within a family of models to cater to different needs:

• TabPFN-3.5: The open-weights base model with 220 million parameters and 8 estimators by default.
• TabPFN-3.5-Fast (alpha): An open-weights version with 84 million parameters, designed to be up to 6 times faster.
• TabPFN-3.5-Plus: Available via API and enterprise licenses, this variant includes native text handling and FP8 attention.
• TabPFN-3.5-Thinking: This version utilizes extra inference compute to enhance performance without relying on LLMs, real data, or search, and is up to 12 times faster than its TabPFN-3 predecessor.

Interestingly, While the open weights are available for research, evaluation, and Kaggle use, production deployments require a commercial license or interaction with Prior Labs’ API.

Why TabPFN-3.5 Matters for the Future of AI

TabPFN-3.5 represents a significant step forward for machine learning on tabular data. Its ability to achieve top-tier performance on complex datasets with minimal setup and no domain-specific tuning opens new possibilities for rapid prototyping, robust deployment, and accessible AI solutions. By simplifying the process of working with tabular data and delivering exceptional results, Prior Labs is paving the way for broader adoption and innovation in this crucial area of artificial intelligence.

Expert Perspective

A practical read on TabPFN-3.5 starts with tabpfn. That is where the earliest effects are likely to show up if this development keeps building.

What happens next will come down to adoption speed, policy response, and execution quality. That combination could make TabPFN-3.5 a meaningful reference point across data.

For decision-makers, the useful lens is not the headline alone but how model changes priorities once organizations have to respond.

Frequently Asked Questions

Why is TabPFN-3.5 important?

A New Era for Tabular Data: Introducing TabPFN-3.5At a glance, The world of artificial intelligence continues to push boundaries, and the latest innovation from Prior Labs marks a significant leap forward for tabular data.

What impact could TabPFN-3.5 have?

Their new foundation model, TabPFN-3.5, has not only demonstrated state-of-the-art performance across numerous benchmarks but has also achieved a remarkable feat: surpassing the winning solution of a prestigious 2015 Kaggle competition, all without any dataset-specific training or tuning.Meanwhile, This release signals a paradigm shift in how we approach machine learning with structured data, offering a powerful, ready-to-use model that can predict on tables in a single forward pass.The Otto Kaggle Challenge: A Historic Feat RevisitedOne of the most compelling demonstrations of TabPFN-3.5’s capabilities comes from its performance on the 2015 Otto Group Product Classification Challenge.

What should readers watch next with TabPFN-3.5?

This Kaggle competition, which attracted over 3,500 teams, tasked participants with classifying products into nine categories based on 93 obfuscated features.

How does this relate to tabpfn?

It connects because the article frames tabpfn as one of the clearest areas where the topic may be felt in practice.

Source: https://www.marktechpost.com/2026/09/15/prior-labs-releases-tabpfn-3-5-a-tabular-foundation-model-that-beats-the-winning-otto-kaggle-solution-with-default-settings/