TabPFN-3, Supporting Up to 1 Million Rows

With the release of TabPFN-3, a state-of-the-art foundation model has emerged that handles tabular data prediction in a single forward pass without training.

Scalability

• Handles up to 1 million rows on a single H100.
• 10x larger scale compared to TabPFN-2.5, with row-chunked inference and reduced KV cache improving practicality.
• Estimated KV cache is about 8GB per 1 million rows per estimator.

Speed

• Inference is 10x to 1000x faster than the previous version.
• SHAP is reported to be 120x faster thanks to KV caching.

Performance

• The API-only Thinking Mode performs additional fitting via test-time compute, scoring over 200 Elo higher than non-TabPFN methods on TabArena.
• In the large-data regime, the gap widens to 420 Elo, with a claimed 93% win rate against classical ML.

Features and Deployment

• Provides a non-parametric retrieval decoder supporting up to 160 classes.
• A bar-distribution head for calibrated quantile regression produces prediction intervals in a single forward pass.
• Claims improvements in time series, interpretability, and relational benchmarks as well.
• Offers three deployment paths: API, enterprise licensing, and open-source weights.