TabPFN-2.5: Advancing the State of the Art in Tabular Foundation Models
Submitted in November 2025. TabPFN-2.5 is the next-generation tabular foundation model supporting up to 50,000 data points and 2,000 features, with a 20x increase in data units over TabPFNv2. On the TabArena benchmark, it significantly outperforms tuned tree models, matching the accuracy of AutoGluon 1.4 (four-hour tuned ensemble). Default TabPFN-2.5 achieves a 100% win rate against default XGBoost on small-to-medium classification datasets, and 87% on larger datasets. A new distillation engine converts the model into compact MLPs or tree ensembles, maintaining most accuracy while drastically reducing latency.
TabPFN-2.5 extends the capability boundary of tabular foundation models to 50,000 samples and 2,000 features, surpassing traditional tree models on standard benchmarks and matching top automated ML systems. Its distillation engine addresses production deployment latency, moving tabular AI from research to large-scale application. This marks a decisive breakthrough for foundation models in the traditionally strong domain of tabular data.
TabPFN-2.5 is based on the Transformer architecture, learning general patterns of tabular data through large-scale pre-training. It supports 50,000 data points and 2,000 features, with a 20x increase in data units over its predecessor. On the TabArena benchmark, the default configuration outperforms tuned XGBoost, LightGBM, etc., matching AutoGluon 1.4 (four-hour tuned ensemble). The distillation engine converts the model into MLPs or tree ensembles, achieving orders-of-magnitude latency reduction while maintaining most accuracy.
TabPFN-2.5 has significant impact on industries reliant on tabular data, such as finance, healthcare, and e-commerce. Its ability to outperform traditional tree models without tuning can greatly lower the barrier to machine learning application. The distillation engine makes it suitable for production environments, potentially replacing traditional gradient boosting trees as the preferred method for tabular data modeling.
Data science teams can evaluate integrating TabPFN-2.5 into existing tabular data modeling pipelines to replace XGBoost/LightGBM. It is recommended to start with small-to-medium datasets, leveraging its zero-tuning capability for rapid validation. For production environments, use the distillation engine to generate compact models to reduce inference costs.
Attention should be paid to TabPFN-2.5's performance on larger datasets (>100,000 samples), as well as its interpretability and feature importance analysis capabilities. The accuracy loss after distillation needs to be evaluated in specific scenarios. Additionally, model updates and continual learning mechanisms will affect its long-term applicability.