Synthetic-Prior In-Context Models
6 entriesModels pretrained on large numbers of synthetic tables, which classify or regress a new dataset in a single forward pass. Descriptions follow each source; they are not endorsements or rankings.
Real-Data and Text-Aware Models
8 entriesModels that use real tables, language models or text encoders, which helps when columns contain free text or meaningful names.
Libraries, APIs and Integrations
6 entriesWays to run these models on your own data.
Benchmarks and Evaluation
5 entriesWhere classification and regression results are compared across many datasets.
Gradient Boosting Baselines
3 entriesThe tree ensembles that tabular foundation models are usually compared against.
Foundational Papers
6 entriesThe research that established in-context learning for tables.
Related Pages
5 entriesRelated directories on this site.
Frequently Asked Questions
What is a tabular foundation model?
It is a model pretrained on a very large number of tables, often synthetic ones, so it can predict the target column of a new table without training from scratch. You give it labelled rows as context and it predicts the unlabelled ones in a forward pass.
Can they handle both classification and regression?
Yes. Models such as TabPFN, TabICL, Mitra and TabFM offer classifier and regressor interfaces. Some models have limits, such as a maximum number of classes or features, so check each model card before choosing.
Do tabular foundation models beat XGBoost or CatBoost?
They often match or beat tuned gradient boosting on small and medium tables, according to benchmarks such as TabArena and the authors' own reports. Gradient boosting remains strong on very large tables, and speed, memory and licences differ, so test both on your data.
How much data do they need?
Very little for training, since there is no gradient update, but they need labelled rows as context. Context size limits differ by model and version, and larger tables usually need a GPU.
Are they free to use commercially?
Not always. Code licences and weight licences can differ, and some weights carry non-commercial terms. Read the licence on each model card and repository before using a model in production.