Prior Labs releases TabPFN-3.5, scaling tabular foundation models to millions of rows
Prior Labs officially released TabPFN-3.5 on Sep 15-16, scaling tabular foundation models from roughly 1,000 rows to over 1 million rows and up to 20,000 features, with no domain-specific training required. According to Prior Labs employee tuanacelik and co-founder FrankRHutter, the new version tops both TabArena and BeyondArena, achieving SOTA in million-row settings; out of the box, without tuning, it reached rank 1 on the classic Kaggle Otto competition leaderboard, which innixma (in a repost) argued is more impressive than any benchmark in the report, indicating untapped signal in the data.
Confirmed
- The release includes multiple variants, among them TabPFN-3.5-Fast with claimed inference speedups of up to 6x
- The new version breaks the iid assumption, supporting time-series data, grouped data, string/text columns, wide tables, and high-cardinality categorical features
- A single model covers both classification and regression with strong uncertainty calibration
- Tops major tabular benchmarks including TabArena and BeyondArena; rank 1 out of the box on Kaggle Otto
Why it matters
- Tabular data is the most common data form in industry; prior tabular foundation models only handled small datasets. TabPFN-3.5 extends coverage to million-row scale and complex data shapes (time series, groups, text columns), substantially closing the gap between training-free foundation models and real industrial scenarios
- Topping a classic Kaggle competition without any tuning demonstrates the potential of pretrained foundation models to replace traditional feature engineering and tuning workflows
2026-09-15 ~ 2026-09-16 · 5 related posts
Primary sources
- Prior Labs ships TabPFN-3.5, SOTA tabular foundation model leading both TabArena and BeyondArena — tuanacelik ·
- Prior Labs releases TabPFN-3.5, a tabular foundation model handling 1M rows, tops benchmarks — tuanacelik ·
- TabPFN-3.5 launches with SOTA on complex tabular data, up to 6x faster inference — FrankRHutter ·
- [source] TabPFN-3.5 launches with SOTA on complex tabular data, up to 6x faster inference — FrankRHutter · 2026-09-15
- [source] Prior Labs ships TabPFN-3.5, SOTA tabular foundation model leading both TabArena and BeyondArena — tuanacelik · 2026-09-16
- TabPFN-3.5 Released, Topping All Relevant Tabular Benchmarks — ph_singer · 2026-09-16
- [source] Prior Labs releases TabPFN-3.5, a tabular foundation model handling 1M rows, tops benchmarks — tuanacelik · 2026-09-16
- TabPFN-3.5 ships, instantly ranking #1 on the classic Kaggle Otto competition — ph_singer · 2026-09-16