TabPFN-3.5: Technical Report

TabPFN-3.5 is a tabular foundation model designed to extend in-context prediction beyond clean i.i.d. datasets to temporal and grouped splits, high-cardinality categorical variables, wide tables, strings, text, images, relational data, and predictive distributions. On TabArena, it ranks 1st of 89 methods, while the broader TabPFN-3.5 family ranks first across seven benchmarks, with mean win rates of 78% against other tabular foundation models and 89% against non-foundation models such as XGBoost, CatBoost, and tuned MLPs. The model supports up to 1M rows and 6K features with 220M parameters, using Fourier value encodings, in-context ECDF features, shared classification and regression representations, and a synthetic prior targeted at grouped, wide, and high-cardinality data. Its transformer width increases from 512 to 1024 dimensions, while grouped-query attention keeps cached prediction latency close to TabPFN-3. TabPFN-3.5-Fast runs up to 3 times faster than TabPFN-3, and TabPFN-3.5-Thinking scales inference-time computation without LLMs or external data. On ScoringBench, TabPFN-3.5 achieves a 2.85 mean rank and improves over TabPFN-3 on 85 of 101 datasets, demonstrating gains not only in point prediction but also in calibrated predictive distributions. The Plus variants add native text and date handling and FP8 attention for deployment-oriented multimodal inference.