Prior Labs Releases TabPFN-3.5: Beats Otto Kaggle
Prior Labs launches TabPFN-3.5, a tabular foundation model beating the 2015 Otto Kaggle winning solution with raw data and default settings.
Stock photo for illustration only, not from the actual event
- Prior Labs has released TabPFN-3.5, an advanced tabular foundation model.
- Scores 0.375 on the private Otto Kaggle leaderboard, outperforming the 2015 winning solution using default settings.
- Pretrained exclusively on synthetic data without prior exposure to Otto or Kaggle datasets.
- Open weights available locally for research and evaluation, with API access for production.
The artificial intelligence community is buzzing as Prior Labs officially rolls out TabPFN-3.5, a cutting-edge foundation model engineered specifically for tabular data. This release has captured the attention of data scientists by demonstrating an exceptional ability to rival legendary competition benchmarks. The model's open weights can be run locally for research and evaluation purposes, while production deployments require Prior Labs' API or a commercial license.
Looking back, the Otto Group Product Classification Challenge took place on Kaggle in 2015, drawing 3,505 teams competing for a $10,000 prize. Entrants were tasked with sorting products into 9 distinct categories using 93 obfuscated count features, evaluated through multi-class log loss where lower scores indicate better performance. The winning solution back then was forged by Gilberto Titericz and Stanislav Semenov, both former world number one Kaggle grandmasters who built a complex multi-layer stack of 36 models leveraging hand-crafted features.
Photo by CHUTTERSNAP / Unsplash
Nick Erickson, co-creator of AutoGluon and an AI researcher at Prior Labs, spent years chasing that benchmark score. According to Erickson, AutoGluon secured rank 23 in its 2020 paper, climbed to rank 14 in 2023, and reached rank 9 in August 2026. The final stretch proved to be the most challenging hurdle; moving from rank 50 down to rank 10 successfully cut log loss from 0.41 to 0.40, but bridging the remaining gap from rank 10 to the winning score of 0.382 required nearly double that effort.
The breakthrough arrives as TabPFN-3.5 achieves an impressive 0.375 score on the private leaderboard. Erickson reports that the model executed on raw data utilizing default settings, completing the task in roughly one minute on an RTX PRO 6000 GPU. Remarkably, the model was pretrained solely on synthetic data, meaning it never encountered the Otto or any other Kaggle dataset beforehand. A fully reproducible Kaggle notebook has also been made publicly available for verification.
The capability of TabPFN-3.5 to surpass a historic Kaggle-winning solution straight out-of-the-box using default settings marks a major milestone for Tabular Foundation Models. Historically, tabular data has remained a challenging domain for standard deep learning architectures compared to gradient boosting frameworks like XGBoost or LightGBM. By leveraging Prior-Data Fitted Networks (PFNs) trained on synthetic data to learn learning algorithms, these models can make accurate predictions on real-world tables instantly, significantly reducing manual feature engineering overhead.
Technical reports further highlight that the model secured first place across multiple benchmarks, including TabArena, BeyondArena, STRABLE, MulTaBench, RelArena-α, TALENT, and ScoringBench. On TabArena, a dynamic benchmark encompassing 51 datasets, the Thinking variant achieves an Elo rating of 1910, while the base model scores 1866, outperforming TabFM+ at 1823. Furthermore, Prior Labs states that the base model beats AutoGluon 1.6 extreme by 130 Elo points in a fraction of the time.
Regarding architectural efficiency, the KV cache remains roughly equivalent in size to TabPFN-3 despite housing approximately four times more parameters. Cached single-row predictions match the speed of TabPFN-3, although on massive training sets, the base model can run up to twice as slow. Enthusiasts can explore the official technical report and details directly through the original publication source.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment