Meet TabPFN-3
The tabular foundation model for state-of-the-art predictions on structured data.
TabPFN-3 by the numbers
93% win rate over classic ML
On TabArena, leading benchmark for structured data tasks
0.2s for predictions on 1M samples
Fast inference for real-world predictive workflows
+420 ELO with thinking mode
TabPFN-3-Plus with Thinking mode beats AutoML in 80% of the cases
Testimonials
TabPFN-3 is a meaningful step beyond prior tabular foundation models. In our benchmarks it has been competitive with strong baselines and, with the right configuration, can outperform even fine-tuned gradient-boosted decision trees. At Affirm, we evaluate frontier techniques like this because the quality of our risk and decisioning models has a direct impact on the experience we deliver to consumers and merchants.
Our collaboration with Prior Labs presents us with an opportunity to test cutting-edge tabular foundation models that have the potential to transform our business processes, forecasting and risk systems.
Hitachi is committed to advancing predictive maintenance across our rail networks. Working with Prior Labs and TabPFN helps us accelerate our use of data for reliability and safety while reducing operational overhead.
Complex lung diseases are one of the most challenging areas in medicine. TabPFN allows us to detect across the broad range of the proteome to deliver actionable clinical innformation with greater efficiency, helping us push the boundaries of liquid biopsy and improve diagnostic accuracy at scale.
%202.svg)
TabPFN presents an exciting opportunity to adopt tabular foundation models and unlock their potential to serve more customers with greater precision.
Production-grade predictions in a single forward pass — typically measured in seconds.
One model for every prediction task
Churn, fraud, pricing, demand forecasting. Serve all use-cases with one model.
Works with your data as is
Missing values, outliers, categoricals, multi-table datasets. Feed raw data and skip feature engineering.
Predictions in seconds, not months
Skip pre-processing and model tuning, and get production-grade predictions from your first predict call.
Built for production-scale datasets
From world-class research to production-ready system. Now handling up to 1M rows natively.
Fast inference
20x faster inference speed than previous model versions and up to 1000x faster inference at scale.
Deploy anywhere.
Keep control of your data.
.avif)
%20copy.avif)
Join the next wave
